Battery state-of-charge estimation method suitable for non-open-loop working state of battery

Abstract

The invention discloses a battery state-of-charge estimation method suitable for a battery non-open-loop working state, and relates to the technical field of batteries. The method sequentially comprises the following steps of: carrying out initial estimation on the state of charge of a battery by utilizing a traditional GRU (Gated Recirculation Unit) algorithm; the relation between the algorithms is obtained by using the result of the GRU algorithm and the input condition of the Kalman filtering algorithm; establishing a battery mathematical model under a fractional order model on the basis of a fractional order theory; introducing an adaptive genetic algorithm (AGA) to obtain parameter identification results of the battery at different temperatures; filtering and correcting the initially estimated state of charge by using a fractional order multi-information adaptive unscented Kalman filtering algorithm; according to the method, the fractional order multi-innovation adaptive Kalman filtering algorithm is added on the basis of a traditional gating cycle unit GRU algorithm, and the influence that a final result is not accurate enough and fluctuates around a real value due to the fact that a single gating cycle unit GRU algorithm cannot filter noise is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of batteries, and in particular to a battery state of charge estimation method suitable for a battery non-open-loop working state. Background Art

[0002] Lithium-ion batteries have become the preferred power source for electric vehicles due to their large capacity, long cycle life, and low memory effect. SOC is one of the most important states that the battery management system (BMS) of electric vehicles needs to supervise and monitor. High-precision and fast-converging SOC state estimation can provide more accurate mileage estimation and expand the available SOC range, avoiding battery loss caused by overcharging and over-discharging, accelerated life decay, combustion, and even explosion.

[0003] In most existing studies, capacitors and inductors are analyzed as integer-order devices. However, in reality, there are no ideal integer-order devices, but parasitic resistance. The commonly used equivalent model is an ideal capacitor, inductor and resistor in series, and this resistor is ESR. Most studies use integer-order models of devices to analyze circuits, which has large errors, so fractional-order theory has been applied in the field of power electronics.

[0004] Since the traditional Kalman filter algorithm estimates and updates the next step based on the estimation result of the previous step, this will have high requirements on the initial value and the estimation state of the previous step. At present, due to the inevitable errors in the battery measurement process, the function cannot converge. Therefore, accurate previous step estimation plays a big role in battery estimation. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned problem. A battery state of charge estimation method suitable for the non-open-loop working state of the battery is designed. A fractional-order multi-innovation adaptive Kalman filter algorithm is added to the traditional gated cyclic unit algorithm, which improves the previous single gated cyclic unit algorithm that cannot filter out noise, resulting in the final result being inaccurate and fluctuating around the true value.

[0006] The object of the present invention is achieved in that:

[0007] A battery state of charge estimation method applicable to a battery non-open-loop working state comprises the following steps:

[0008] Step a, using the traditional gated recurrent unit GRU algorithm to make an initial estimate of the battery state of charge;

[0009] Step b, using the result of the traditional gated recurrent unit GRU algorithm and the input condition of the Kalman filter algorithm to obtain the connection between the algorithms;

[0010] Step c, establishing a battery mathematical model under a fractional order model based on fractional order theory;

[0011] Step d, introducing an adaptive genetic algorithm (AGA) to obtain parameter identification results of batteries at different temperatures;

[0012] Step e: Filter and correct the initially estimated state of charge using a fractional-order multi-innovation adaptive unscented Kalman filter algorithm.

[0013] The above-mentioned method for estimating the state of charge of a battery applicable to a non-open-loop working state of the battery, the method adopted in step a is an improvement of the neural network algorithm LSTM, and the steps are:

[0014] Step a1: using a gated cycle unit algorithm to preliminarily identify the SOC of the lithium-ion battery;

[0015] Step a2, establishing a relationship between the gated recurrent unit algorithm and the fractional-order multi-innovation adaptive unscented Kalman filter algorithm;

[0016] Step a3, creating a battery fractional-order model based on the relationship between the integer-order theorem and the fractional order;

[0017] Step a4, using an adaptive genetic algorithm to identify parameters of a fractional-order model of a lithium-ion battery;

[0018] Step a5: Use a fractional-order multi-innovation adaptive unscented Kalman filter algorithm to eliminate filtering noise from the initial estimated SOC and correct the estimated value.

[0019] In the step a, the neural network algorithm LSTM is improved, a network structure diagram of the gated recurrent unit GRU model is established, and the program of the gated recurrent unit GRU algorithm is obtained according to the structure diagram, and the initial state of charge of the battery is estimated. The update gate and reset gate formulas are expressed as shown in the following formula:

[0020] Z t =σ(W z [h t-1 ,x t ])

[0021] r t =σ(W r [h t-1 ,x t ])

[0022] Input x t The terminal voltage U obtained in the initial experiment is the current state t and loop current I t ,h t-1 Indicates the hidden state of the previous moment, which is h tThe previous state value is used to predict the result of the next state. The calculation weights of the update gate and the reset gate are respectively determined by the weight parameters W r and W z control, represents the pre-estimation of the hidden layer data, The candidate set is obtained by combining the current input x t and input reset gate mechanism, The expression is as follows:

[0023]

[0024] After the tanh activation function, is normalized to the interval [-1,1] and then updated with the gate Z t Multiply them together to get the data h passed to the hidden layer at the next moment t ,h t The expression is as follows:

[0025]

[0026] The output expression y of the GRU neural network t , as shown below:

[0027] y t =σ(W o *h t +b y )

[0028] Where W o and b y Represents the weights and thresholds of the output layer.

[0029] The above-mentioned method for estimating the state of charge of a battery applicable to a non-open-loop working state of the battery, in the step b, finds the connection between the gated recurrent unit GRU algorithm and the Kalman filter algorithm according to the definition of the Kalman filter algorithm, in the gated recurrent unit GRU, the input parameter is the voltage and current of the battery currently being tested, that is: x = [UI], and the output parameter is the estimated current battery state of charge SOC;

[0030] In the Kalman filter algorithm, the input is the current capacitor voltage U1U2 and the current initial estimated state of charge of the battery The input parameters in the prior state variables and their variances of the fractional-order model of the updating battery are:

[0031]

[0032] in, and is the estimated result of the gated recurrent unit GRU, and the charge state estimated by the gated recurrent unit GRU is used as

[0033] The above-mentioned battery state of charge estimation method applicable to the non-open-loop working state of the battery, the specific fractional order theory in step c is GL, and the specific expression is:

[0034] Step c1, using the characteristics of battery capacitance and resistance to obtain a fractional-order model of the battery, where Z represents the fractional-order resistance, C f represents the fractional capacitance,

[0035]

[0036] The fractional calculus formula is given by:

[0037]

[0038] When the value of a is different, the type of function is also different. When a>0, the function is differential; when a=0, the function is the original function; when a<0, the function is integral, as follows:

[0039]

[0040] in, is the differential operator with respect to variable t, a is the order of the system, h is the sampling interval, L is the memory or window length, and the factor and is the Newton binomial coefficient, so we have j=1,2,……,k;

[0041] Step c2: Replace the capacitor of the second-order RC model with a fractional-order capacitor, and establish the state space equation of the fractional-order RC model as follows:

[0042]

[0043] Step c3, combine the above formula with the GRU initial estimated state value to obtain a discrete fractional second-order RC model:

[0044]

[0045] U k =C k x k -I k R0+E+v k

[0046] in, C k=[-1,-1,0], state vector x k =[U1(k),U2(k),SOC(k)] T , and fractional order factors w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k .

[0047] The above-mentioned method for estimating the battery state of charge applicable to the non-open-loop working state of the battery, the algorithm used in the step d is a genetic algorithm, and the specific formula is:

[0048] The adaptive genetic algorithm is used to obtain the parameter identification of the battery fractional-order model. The objective function of the intelligent algorithm in parameter identification is the measured output Y r (j) and the output Y predicted by the model e The difference of (j) is the smallest, and its mathematical model is expressed as:

[0049]

[0050] Where M is the length of the data set, and the fitness function is expressed as:

[0051]

[0052] The crossover probability and mutation probability of the AGA algorithm are expressed as:

[0053]

[0054] For individuals with J(i)<J(max), we have:

[0055] CrossoverRat(i)=K2

[0056] MutationRat(i)=K4

[0057] Among them, K1, K2, K3 and K4 are all ≤1.

[0058] In step e, the relationship in step b is actually applied to combine the result obtained by the gated recurrent unit GRU with the fractional-order multi-innovation adaptive unscented Kalman filter algorithm. The specific combination process formula is:

[0059] Fractional-order multi-innovation adaptive unscented Kalman filter correction algorithm program:

[0060] Nonlinear state-space systems:

[0061]

[0062] Among them, w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k .

[0063] Step e comprises the following steps:

[0064] Step e1, data initialization is performed according to the following algorithm:

[0065]

[0066] in, For the initial input data, is the initial error covariance, is the process noise covariance, is the measurement noise covariance;

[0067] Step e2 comprises the following steps:

[0068] Step e21, according to the state space equation of the fractional order model, sample the particles and sigma them:

[0069]

[0070] Step e22, calculate weight:

[0071]

[0072] Where L represents the dimension of the window, where x k-1 is the model state variable at time k-1, λ is the proportional factor, X i is the sigma point set, P l-1 is the state estimation covariance,

[0073] Step e23: Update the prior state variables and their variances of the fractional-order model according to the pre-estimation:

[0074]

[0075] in, represents the state estimator of the k-1th step, Indicates that it contains the state estimate of the previous step and the current u k The state update value, P xx is the update of the state variance matrix, is the prior state variable, Q k-1 is the process noise covariance at time k-1;

[0076] Step e3: Introduce the innovation matrix and update the multiple innovation observation variables and their covariances:

[0077]

[0078] in represents the state estimate of the k-th step, represents multiple innovative observation variables, P yy is the update of the variance of the observed variable;

[0079] Step e4: Update the joint covariance and Kalman matrix gain at time K+1:

[0080]

[0081] Step e5: Introduce the adaptive part to update the noise covariance:

[0082]

[0083] Real-time variance estimation with multiple innovations

[0084]

[0085] Adaptive noise covariance update:

[0086]

[0087] H represents the noise covariance matrix, d k represents the filtering error, W represents the length of the innovation error window;

[0088] Step e6: Update the posterior state value and its covariance matrix:

[0089] y k =U k

[0090]

[0091] P k =P xx -K p,k P xy T

[0092] Among them, K p,k is the Kalman gain, P xy is the joint covariance, is the prior state variable.

[0093] Compared with the prior art, the battery state of charge estimation method of the present invention has the following beneficial effects:

[0094] First, on the basis of the traditional gated recurrent unit GRU algorithm, a fractional-order multi-information adaptive Kalman filter algorithm FOMIAUKF is added, which improves the previous single gated recurrent unit GRU algorithm that cannot filter out noise, resulting in the final result being inaccurate and fluctuating around the true value. In addition, due to the addition of the gated recurrent unit GRU algorithm, the initial estimate of the filtering algorithm is not far from the true value, which solves the problem of being unable to accurately predict the range of noise errors, resulting in a large gap between the final prediction result and the true value and the appearance of parallel curves. Compared with the traditional gated recurrent unit GRU algorithm, the error of the battery under DST conditions at different temperatures can be reduced by about 30%, greatly increasing the robustness.

[0095] Second, the gated recurrent unit GRU algorithm can be used to preliminarily estimate the battery state of charge to compensate for the problem of discrete and excessive estimation errors caused by inaccurate initial state estimation in the FOMIAUKF algorithm. The present invention first uses the gated recurrent unit GRU algorithm to obtain an estimated value of the initial SOC, and then, based on the traditional integer-order model, uses the characteristics of capacitors and inductors to establish a fractional-order model on the basis of the second-order RC; then introduces an adaptive genetic AGA algorithm to realize parameter identification of the model, and uses the fractional-order multi-information adaptive unscented Kalman filter algorithm FOMIAUKF under the obtained optimal parameters to verify the accuracy of online estimation of lithium-ion batteries under HPPC and DST conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 It is an overall schematic diagram of the estimation method of the present invention;

[0097] Figure 2 It is a flow chart of the battery state of charge estimation method based on GRU-FOMIAUKF combination according to the present invention;

[0098] Figure 3 This is a diagram of the structure model of the gated recurrent unit (GRU) neural network described in the present invention;

[0099] Figure 4 is a flow chart of the adaptive genetic algorithm AGA of the present invention;

[0100] Figure 5 It is a flowchart of the GRU-FOMIAUK process algorithm described in the present invention;

[0101] Figure 6 It is the lithium ion battery EIS and the integer-order and fractional-order equivalent circuit model diagram of the lithium ion battery of the present invention;

[0102] Figure 7 This is a graph showing the SOC estimation result of DST by GRU-FOMIAUKF at 25°C according to the present invention;

[0103] Figure 8 This is a graph showing the SOC estimation results of HPPC using GRU-FOMIAUKF at 25°C as described in the present invention. DETAILED DESCRIPTION

[0104] The specific implementation modes of the present invention will be further described in detail below with reference to the accompanying drawings. Specific implementation method 1

[0106] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation mode is as follows: Figure 1 As shown, the following steps are included:

[0107] Step a, using the traditional gated recurrent unit GRU algorithm to make an initial estimate of the battery state of charge;

[0108] Step b, using the result of the traditional gated recurrent unit GRU algorithm and the input condition of the Kalman filter algorithm to obtain the connection between the algorithms;

[0109] Step c, establishing a battery mathematical model under a fractional order model based on fractional order theory;

[0110] Step d, introducing an adaptive genetic algorithm (AGA) to obtain parameter identification results of batteries at different temperatures;

[0111] Step e: Filter and correct the initially estimated state of charge using a fractional-order multi-innovation adaptive unscented Kalman filter algorithm. Specific implementation method 2

[0113] The battery state of charge estimation method applicable to the battery non-open-loop working state in this specific implementation is as follows: Figure 2 As shown, based on the specific implementation mode 1, further definition is made:

[0114] The method adopted in step a is an improvement of the neural network algorithm LSTM, and the steps are:

[0115] Step a1: using a gated cycle unit algorithm to preliminarily identify the SOC of the lithium-ion battery;

[0116] Step a2, establishing a relationship between the gated recurrent unit algorithm and the fractional-order multi-innovation adaptive unscented Kalman filter algorithm;

[0117] Step a3, creating a battery fractional-order model based on the relationship between the integer-order theorem and the fractional order;

[0118] Step a4, using an adaptive genetic algorithm to identify parameters of a fractional-order model of a lithium-ion battery;

[0119] Step a5: Use a fractional-order multi-innovation adaptive unscented Kalman filter algorithm to eliminate filtering noise from the initial estimated SOC and correct the estimated value. Specific implementation method three

[0121] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation is based on the second specific implementation mode, and further defines:

[0122] In the step a, the neural network algorithm LSTM is improved, a network structure diagram of the gated recurrent unit GRU model is established, and the program of the gated recurrent unit GRU algorithm is obtained according to the structure diagram, the initial state of charge result of the battery is estimated, and the gated recurrent unit GRU algorithm is obtained according to the improvement of the neural network algorithm LSTM in the hidden layer. The update gate and reset gate formulas are expressed as shown in the following formula:

[0123] Z t =σ(W z [h t-1 ,x t ])

[0124] r t =σ(W r [h t-1 ,x t ])

[0125] Input x t The terminal voltage U obtained in the initial experiment is the current state t and loop current I t ,h t-1 Indicates the hidden state of the previous moment, which is h t The previous state value is used to predict the result of the next state. The calculation weights of the update gate and the reset gate are respectively determined by the weight parameters W r and W z The opening and closing degree of these two doors is controlled by the Sigmoid activation function σ, which has a value range of [0,1] and is used to adjust the opening and closing degree of and. represents the pre-estimation of the hidden layer data, The candidate set is obtained by combining the current input x t and input reset gate mechanism, The expression is as follows:

[0126]

[0127] After the tanh activation function, is normalized to the interval [-1,1] and then updated with the gate Z t Multiply them together to get the data h passed to the hidden layer at the next moment t,h t The expression is as follows:

[0128]

[0129] according to Figure 3 The gated recurrent unit GRU neural network structure model diagram is shown, and the gated recurrent unit GRU battery state of charge estimation method program is constructed. Figure 3 The final result in is y t , the output expression y of the GRU neural network t , as shown below:

[0130] y t =σ(W o *h t +b y )

[0131] Where W o and b y Represents the weights and thresholds of the output layer.

[0132] The LOESS algorithm is introduced to initially obtain a result close to the true value. Since the gated recurrent unit GRU algorithm is not filtered, the SOC estimate of the lithium-ion battery will fluctuate around the true value. The introduction of the LOESS algorithm will make the result of the gated recurrent unit GRU smoother and closer to the true lithium-ion battery SOC estimate. However, compared with the true value, it lacks two important parts, filtering and correction. The FOMIAUKF algorithm will be used below to further correct and filter the state of charge. Specific implementation method four

[0134] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation is further defined on the basis of the specific implementation method 1: in the step b, the connection between the gated recurrent unit GRU algorithm and the Kalman filter algorithm is found according to the definition of the Kalman filter algorithm. In the gated recurrent unit GRU, the input parameter is the voltage and current of the battery currently tested, that is, x=[UI], and the output parameter is the estimated current battery state of charge SOC;

[0135] In the Kalman filter algorithm, the input is the current capacitor voltage U1U2 and the current initial estimated state of charge of the battery The input parameters in the prior state variables and their variances of the fractional-order model of the updating battery are:

[0136]

[0137] in, and is the estimated result of the gated recurrent unit GRU, and the charge state estimated by the gated recurrent unit GRU is used as Specific implementation method five

[0139] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation is further defined on the basis of the specific implementation mode 1: the specific fractional order theory in step c is GL, and the specific expression is:

[0140] Step c1, using the characteristics of battery capacitance and resistance to obtain a fractional-order model of the battery, where Z represents the fractional-order resistance, C f represents the fractional capacitance,

[0141]

[0142] The fractional calculus formula is given by:

[0143]

[0144] When the value of a is different, the type of function is also different. When a>0, the function is differential; when a=0, the function is the original function; when a<0, the function is integral, as follows:

[0145]

[0146] in, is the differential operator with respect to the variable t, a is the order of the system, h represents the sampling interval, L is the memory or window length, and the factor and is the Newton binomial coefficient, so we have j=1,2,……,k;

[0147] Step c2: Replace the capacitor of the second-order RC model with a fractional-order capacitor, and establish the state space equation of the fractional-order RC model as follows:

[0148]

[0149] Step c3, combine the above formula with the GRU initial estimated state value to obtain a discrete fractional second-order RC model:

[0150]

[0151] U k =C k x k -I k R0+E+v k

[0152] in, C k =[-1,-1,0], state vector x k =[U1(k),U2(k),SOC(k)] T , and fractional order factors w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k .

[0153] The data is obtained by using the dynamic stress test method DST and the hybrid pulse discharge test HPPC test of NCR18650 battery at 15℃, 25℃, and 35℃. The voltage and current obtained from the experimental results are as follows Figure 5-8 As shown, the open circuit voltage of the hybrid pulse discharge experiment obtained in the experiment and the SOC experimental data obtained at present are fitted with a polynomial to obtain the OCV-SOC fitting curve. By analyzing the fitting results of different orders, the 6th order result is the best.

[0154] The electrochemical impedance spectroscopy (EIS) experiment of lithium-ion batteries can be used to study the electrochemical process inside the battery, the performance of electrode materials, and the battery kinetic characteristics, providing a basis for establishing fractional-order models. The results of the EIS experiment can be represented by the Nyquist curve. The Nyquist curve of a lithium-ion battery under the action of an AC signal is shown in the figure below. Figure 4 shown. Figure 4 It can be divided into three parts: low frequency, medium frequency and high frequency, which respectively reflect the electrochemical and kinetic processes in different states inside the battery. The low frequency band is a slant line, which is related to the impedance of the lithium ion diffusion process; the medium frequency band is a semi-ellipse, which is related to the impedance of the charge transfer process and the double-layer capacitance effect; the high frequency band is a vertical line, which is related to the inductance characteristics of the battery.

[0155] Since the integer order model is quite different from the fractional order model, the least square method of forgetting factor cannot be used for identification. In this specific implementation, an adaptive genetic algorithm is used for identification. The specific flow chart is as follows: Figure 5 As shown in the figure, the adaptive genetic algorithm is an improvement on the genetic algorithm. The specific formula is as follows: Specific implementation method 6

[0157] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific embodiment is based on the specific embodiment 1, and further defines that the algorithm used in the step d is a genetic algorithm. According to the genetic algorithm structure diagram, the adaptive genetic algorithm is used to obtain the parameter identification of the battery fractional-order model. The objective function of the intelligent algorithm in the parameter identification is the output Y obtained by measurement. r (j) and the output Y predicted by the model e The difference of (j) is the smallest, and its mathematical model is expressed as:

[0158]

[0159] Where M is the length of the data set, and the fitness function is expressed as:

[0160]

[0161] The crossover probability and mutation probability of the AGA algorithm are expressed as:

[0162]

[0163] For individuals with J(i)<J(max), we have:

[0164] CrossoverRat(i)=K2

[0165] MutationRat(i)=K4

[0166] Among them, K1, K2, K3 and K4 are all ≤1.

[0167] The specific recognition results are shown in Table 1 below. The recognition results are applied to the battery FOMIAUKF framework below. Figure 4 shown.

[0168] Table 1 Parameter identification results of battery fractional-order model at 25℃

[0169]

[0170] The implementation of the traditional unscented Kalman UKF filtering algorithm is based on an ideal assumption about noise, and usually the noise of the system model is assumed to be Gaussian white noise that obeys a normal distribution. However, the system model noise will change over time and is not completely known. Therefore, in this specific implementation, on the basis of FOM, the adaptive covariance algorithm is combined with the multi-information method to propose the FOMIAUKF algorithm, that is, the fractional-order multi-information adaptive unscented Kalman filtering algorithm. The adaptive covariance algorithm reduces the estimation error of SOC according to the formula Rk in real time during the SOC estimation process. Specific implementation method seven

[0172] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation is further defined on the basis of the specific implementation method 6: in the step e, the relationship in the above step b is actually applied to combine the result obtained by the gated recurrent unit GRU with the fractional-order multi-innovation adaptive unscented Kalman filter algorithm. The specific combination process formula is:

[0173] Fractional-order multi-innovation adaptive unscented Kalman filter correction algorithm program:

[0174] Nonlinear state-space systems:

[0175]

[0176] Among them, w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k . Specific implementation method eight

[0178] The battery state of charge estimation method applicable to the non-open-loop working state of the battery in this specific implementation is further defined on the basis of the specific implementation method 7:

[0179] Step e comprises the following steps:

[0180] Step e1, data initialization is performed according to the following algorithm:

[0181]

[0182] in, For the initial input data, is the initial error covariance, is the process noise covariance, is the measurement noise covariance;

[0183] Step e2 comprises the following steps:

[0184] Step e21, according to the state space equation of the fractional order model, sample the particles and sigma them:

[0185]

[0186] Step e22, calculate weight:

[0187]

[0188] Where L represents the dimension of the window, where x k-1 is the model state variable at time k-1, λ is the proportional factor, Xi is the sigma point set, P l-1 is the state estimation covariance,

[0189] Step e23: Update the prior state variables and their variances of the fractional-order model according to the pre-estimation:

[0190]

[0191] in, represents the state estimator of the k-1th step, Indicates that it contains the state estimate of the previous step and the current u k The state update value, P xx is the update of the state variance matrix, is the prior state variable, Q k-1 is the process noise covariance at time k-1;

[0192] Step e3: Introduce the innovation matrix and update the multiple innovation observation variables and their covariances:

[0193]

[0194] in represents the state estimate of the k-th step, represents multiple innovative observation variables, P yy is the update of the variance of the observed variable;

[0195] Step e4: Update the joint covariance and Kalman matrix gain at time K+1:

[0196]

[0197] Step e5: Introduce the adaptive part to update the noise covariance:

[0198]

[0199] Real-time variance estimation with multiple innovations

[0200]

[0201] Adaptive noise covariance update:

[0202]

[0203] H represents the noise covariance matrix, d k represents the filtering error, W represents the length of the innovation error window;

[0204] Step e6: Update the posterior state value and its covariance matrix:

[0205] y k =Uk

[0206]

[0207] P k =P xx -K p,k P xy T

[0208] Among them, K p,k is the Kalman gain, P xy is the joint covariance, is the prior state variable.

[0209] The experimental results are as follows Figure 7 and Figure 8 As shown, from Figure 7 As can be seen from Table 2, at 25°C, compared with the single gated cyclic unit algorithm, the gap between the estimated value and the true value of the battery state of charge has been improved after the introduction of the filtering algorithm. It can also be seen from the table that the MAE, RMSE, and MSE of the gated cyclic unit algorithm combined with the fractional-order multi-innovation adaptive unscented Kalman filter algorithm have been greatly improved, among which the RMSE and MSE under the DST condition have been improved by 50%. Figure 8 It can be seen from Table 3 that under HPPC conditions, RMSE and MSE are improved by nearly 15%. The experimental results show that the algorithm can show high robustness under different temperatures and different conditions.

[0210] Table 2 is a comparison table of performance parameters of SOC estimation under 25 degrees Celsius DST conditions of the present invention

[0211]

[0212] Table 3 is a comparison table of performance parameters of SOC estimation under 25 degrees Celsius HPPC working conditions of the present invention

[0213]

[0214] It should be noted that the above specific embodiments are only illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above specific embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A battery state of charge estimation method suitable for a battery non-open-loop working state, characterized in that: The steps include: Step a, using the traditional gated recurrent unit GRU algorithm to make an initial estimate of the battery state of charge; Step b, using the result of the traditional gated recurrent unit GRU algorithm and the input condition of the Kalman filter algorithm to obtain the connection between the algorithms; Step c, establishing a battery mathematical model under a fractional order model based on fractional order theory; Step d, introducing an adaptive genetic algorithm (AGA) to obtain parameter identification results of batteries at different temperatures; Step e: Filter and correct the initially estimated state of charge using a fractional-order multi-innovation adaptive unscented Kalman filter algorithm.

2. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 1, characterized in that: The method adopted in step a is an improvement of the neural network algorithm LSTM, and the steps are: Step a1: using a gated cycle unit algorithm to preliminarily identify the SOC of the lithium-ion battery; Step a2, establishing a relationship between the gated recurrent unit algorithm and the fractional-order multi-innovation adaptive unscented Kalman filter algorithm; Step a3, creating a battery fractional-order model based on the relationship between the integer-order theorem and the fractional order; Step a4, using an adaptive genetic algorithm to identify parameters of a fractional-order model of a lithium-ion battery; Step a5: Use a fractional-order multi-innovation adaptive unscented Kalman filter algorithm to eliminate filtering noise from the initial estimated SOC and correct the estimated value.

3. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 2, characterized in that: In the step a, the neural network algorithm LSTM is improved, a network structure diagram of the gated recurrent unit GRU model is established, and the program of the gated recurrent unit GRU algorithm is obtained according to the structure diagram, and the initial state of charge of the battery is estimated. The update gate and reset gate formulas are expressed as shown in the following formula: Z t =σ(W z [h t-1 ,x t ]) r t =σ(W r [h t-1 ,x t ]) Input x t The terminal voltage U obtained in the initial experiment is the current state t and loop current I t ,h t-1 Indicates the hidden state of the previous moment, which is h t The previous state value is used to predict the result of the next state. The calculation weights of the update gate and the reset gate are respectively determined by the weight parameters W r and W z control, represents the pre-estimation of the hidden layer data, The candidate set is obtained by combining the current input x t and input reset gate mechanism, The expression is as follows: After the tanh activation function, is normalized to the interval [-1,1] and then updated with the gate Z t Multiply them together to get the data h passed to the hidden layer at the next moment t ,h t The expression is as follows: The output expression y of the GRU neural network t , as shown below: y t =σ(W o *h t +b y ) Where W o and b y Represents the weights and thresholds of the output layer.

4. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 1, characterized in that: In the step b, the connection between the gated recurrent unit GRU algorithm and the Kalman filter algorithm is found according to the definition of the Kalman filter algorithm. In the gated recurrent unit GRU, the input parameters are the voltage and current of the battery being tested, that is, x=[UI], and the output parameter is the estimated current battery state of charge SOC; In the Kalman filter algorithm, the input is the current capacitor voltage U1U2 and the current initial estimated state of charge of the battery The input parameters in the prior state variables and their variances of the fractional-order model of the updating battery are: in, and is the estimated result of the gated recurrent unit GRU, and the charge state estimated by the gated recurrent unit GRU is used as 5. The method for estimating the state of charge of a battery applicable to a non-open-loop working state of the battery according to claim 1, characterized in that: The specific fractional order theory in step c is GL, and the specific expression is: Step c1, using the characteristics of battery capacitance and resistance to obtain a fractional-order model of the battery, where Z represents the fractional-order resistance, C f represents the fractional capacitance, The fractional calculus formula is given by: When the value of a is different, the type of function is also different. When a>0, the function is differential; when a=0, the function is the original function; when a<0, the function is integral, as follows: in, is the differential operator with respect to the variable t, a is the order of the system, h represents the sampling interval, L is the memory or window length, and the factor and is the Newton binomial coefficient, so we have Step c2: Replace the capacitor of the second-order RC model with a fractional-order capacitor, and establish the state space equation of the fractional-order RC model as follows: Step c3, combine the above formula with the GRU initial estimated state value to obtain a discrete fractional second-order RC model: in, C k =[-1,-1,0], state vector x k =[U1(k),U2(k),SOC(k)] T , and fractional order factors w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k .

6. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 1, characterized in that: The algorithm used in step d is a genetic algorithm, and the specific formula is: The adaptive genetic algorithm is used to obtain the parameter identification of the battery fractional-order model. The objective function of the intelligent algorithm in parameter identification is the measured output Y r (j) and the output Y predicted by the model e The difference of (j) is the smallest, and its mathematical model is expressed as: Where M is the length of the data set, and the fitness function is expressed as: The crossover probability and mutation probability of the AGA algorithm are expressed as: For individuals with J(i)<J(max), we have: CrossoverRat(i)=K2 MutationRat(i)=K4 Among them, K1, K2, K3 and K4 are all ≤1.

7. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 6, characterized in that: In step e, the relationship in step b is actually applied to combine the result obtained by the gated recurrent unit GRU with the fractional-order multi-innovation adaptive unscented Kalman filter algorithm. The specific combination process formula is: Fractional-order multi-innovation adaptive unscented Kalman filter correction algorithm program: Nonlinear state-space systems: Among them, w k is the system noise, v k is the measurement noise, which obeys the normal distribution and is an independent zero-mean Gaussian white noise. The covariance matrices are Q k and R k .

8. A battery state of charge estimation method applicable to a battery non-open-loop working state according to claim 7, characterized in that: Step e comprises the following steps: Step e1, data initialization is performed according to the following algorithm: in, For the initial input data, is the initial error covariance, is the process noise covariance, is the measurement noise covariance; Step e2 comprises the following steps: Step e21, according to the state space equation of the fractional order model, sample the particles and sigma them: Step e22, calculate weight: Where L represents the dimension of the window, where x k-1 is the model state variable at time k-1, λ is the proportional factor, X i is the sigma point set, P l-1 is the state estimation covariance, Step e23: Update the prior state variables and their variances of the fractional-order model according to the pre-estimation: in, represents the state estimator of the k-1th step, Indicates that it contains the state estimate of the previous step and the current u k The state update value, P xx is the update of the state variance matrix, is the prior state variable, Q k-1 is the process noise covariance at time k-1; Step e3: Introduce the innovation matrix and update the multiple innovation observation variables and their covariances: in represents the state estimate of the k-th step, represents multiple innovative observation variables, P yy is the update of the variance of the observed variable; Step e4: Update the joint covariance and Kalman matrix gain at time K+1: Step e5: Introduce the adaptive part to update the noise covariance: Real-time variance estimation with multiple innovations Adaptive noise covariance update: H represents the noise covariance matrix, d k represents the filtering error, W represents the innovation error window length; step e6, update the posterior state value and its covariance matrix: y k =U k P k =P xx -K p,k P xy T Among them, K p,k is the Kalman gain, P xy is the joint covariance, is the prior state variable.

Images