An online lithium battery state of charge estimation method based on state estimation

Abstract

The application discloses an online lithium battery state of charge estimation method based on state estimation, obtains an OCV-SOC characteristic curve by a small current discharge method and a hybrid pulse power performance test method through a second-order RC equivalent circuit model, and then obtains an optimized OCV-SOC curve through a differential evolution algorithm. Parameters are identified online through an adaptive fading factor recursive least square method (AFRLS), the parameters of an unscented Kalman filter observer are updated, a timing model gated recurrent unit network (GRU) adaptive interference noise is introduced, online SOC prediction of a lithium battery is realized as output, finally, the predicted SOC is combined with the OCV-SOC curve to predict the terminal voltage, and the predicted terminal voltage is compared with the measured terminal voltage to form feedback and update; the application can adapt to various working conditions, has low computational complexity, high accuracy, and the accuracy is significantly improved compared with a traditional extended Kalman model and an unscented Kalman model.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, specifically to an online lithium battery state-of-charge estimation method based on state estimation. Background Technology

[0002] Sustainable energy utilization has been a crucial research direction for countries worldwide seeking energy structure transformation, and energy storage technology is one of the key means to promote efficient, clean, and electrified energy use. Lithium-ion batteries, as the most widely used type of battery, face increasingly prominent issues regarding safety and energy storage capacity control. The state of charge (SOC) of a lithium-ion battery, as a parameter reflecting its remaining energy, plays a vital role in addressing these issues. Accurate estimation of SOC is essential to ensure long-term stable operation of lithium-ion batteries, reduce safety hazards, and increase energy storage capacity. Since the state of charge cannot be directly measured, various SOC estimation methods exist.

[0003] State estimation is a closed-loop estimation method based on the battery's equivalent circuit. It involves establishing a pre-built equivalent circuit or electrochemical model to find the nonlinear dynamic relationship between voltage, current, and state of charge (SOC), constructing a state-space equation, and then using an observer to estimate the SOC in this model, such as an extended Kalman filter, an unscented Kalman filter, or a sliding mode observer. Open-loop estimation methods can be implemented using open-circuit voltage-to-SOC correspondence tables or ampere-hour integration methods. Ampere-hour integration calculates the change in SOC by integrating the battery's operating current per unit time. This method is computationally inexpensive and simple to implement, and is widely used in various fields. However, the initial SOC estimate and the cumulative error of the current often introduce drift bias into the final SOC estimate. While estimating the SOC by looking up the open-circuit voltage-to-SOC correspondence table can significantly improve the accuracy of the estimation, in practical applications, the battery needs to be left to stand for a period of time before the open-circuit voltage can be measured, which is not suitable for engineering applications. Data-driven estimation methods estimate the SOC through nonlinear fitting based on measurable parameters such as battery operating voltage, current, and temperature. These methods rely excessively on the amount of existing data; the wider the data coverage, the higher the accuracy. However, they also place extremely high demands on the performance of the battery management system's processor in engineering applications. Kalman filtering (KF) can model and estimate linear systems. To suit nonlinear battery systems, the Extended Kalman Filter (EKF) was developed. EKF performs a first-order Taylor expansion of the nonlinear system and its state transition equations, resulting in insufficient accuracy. Furthermore, patent application CN202110329255.2, "A Method for Estimating the State of Charge of Lithium-ion Batteries Based on Characteristic Curve Optimization," uses recursive least squares (FFRLS) with a forgetting factor for identifying second-order battery model parameters. Because the forgetting factor is fixed, it lacks adaptability to changing system states. Therefore, further improvements to existing technologies are needed. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an online lithium battery state of charge estimation method based on state estimation, so as to predict the online state of charge (SOC) of lithium batteries based on the state estimation method.

[0005] To address the aforementioned technical problems, this invention provides an online lithium battery state-of-charge estimation method based on state estimation. First, a discrete OCV-SOC characteristic curve L1 is obtained using a hybrid pulse power performance testing method. Then, a continuous OCV-SOC characteristic curve L2 is obtained using a small current discharge method. Finally, an optimized continuous OCV-SOC characteristic curve L3 is obtained based on a differential evolution method. The following steps are then performed:

[0006] S01. Using the characteristic curve L3 as the standard, the parameters of the second-order RC battery equivalent circuit model are identified by combining the adaptive attenuation factor recursive least squares method (AFRLS). When used for the first time, the open-circuit voltage OCV when the battery is at rest is input. Subsequent inputs are the open-circuit voltage OCV calculated based on the battery voltage, current and the battery SOC predicted at the previous moment according to the optimized continuous OCV-SOC characteristic curve L3. The output is the parameter identification result of the second-order RC battery equivalent circuit model, including the parametric ohmic resistance R0 of the second-order RC battery equivalent circuit model of the characteristic curve L3 and R1, C1 and R2, C2 in the second-order RC loop.

[0007] S02: Based on the unscented Kalman UKF, introduce the GRU forget gate to establish an unscented Kalman state observer RGUKF model optimized by the gated network. Combine the second-order RC battery equivalent circuit model in step S01, input the measured voltage and current data of the lithium battery under different working conditions collected in real time and the battery SOC predicted at the previous moment, and predict the SOC value at the current moment.

[0008] S03: Convert the SOC value at the current moment to the corresponding open-circuit voltage OCV based on the characteristic curve L3 of the optimized continuous OCV-SOC, and use it as the input for the next moment. Correct the parameters of the adaptive attenuation factor recursive least squares method (AFRLS) and the unscented Kalman state observer RGUKF based on the gated network optimization, and update the model.

[0009] As an improvement to the lithium battery state-of-charge estimation method based on the state estimation method of this invention:

[0010] The specific steps of S01 are as follows:

[0011] S0101. Update the solution using Recursive Least Squares (RLS).

[0012] make The recursive relation is obtained as follows:

[0013]

[0014] in, This is the optimal parameter vector solution at time n;

[0015] q n =[-f(n-1),-f(n-2),…,-f(nm),x(n-1),x(n-2),…,x(nm)] (14)

[0016]

[0017] S0102. Introduce an adaptive attenuation factor c in the recursive least squares (RLS) method. n

[0018]

[0019] c n f is the true value of the system observation at time n. i Compared with the observed estimates The squared error value, where M is the window size for selecting the calculation sequence;

[0020] Parameter vector solution in Adaptive attenuation factor recursive least squares (AFRLS) for:

[0021]

[0022] S0103. The mathematical form of the second-order equivalent model of the battery established by the adaptive attenuation factor recursive least squares method (AFRLS) is given by:

[0023] Define d0 = Δt 2 +wΔt+q, has

[0024]

[0025] w=-d0(d1+2d2) / Δt

[0026] e = d0(d3 + d4 + d5) / Δt 2

[0027] r = -d0(d4 + 2d5) / Δt

[0028] R0 = d5 / d2

[0029] in,

[0030]

[0031] q=τ1τ2, w=τ1+τ2, e=R1+R2+R0, r=R1τ2+R2τ1+R0(τ1+τ2),

[0032] R0 is the internal resistance of the battery point. The R1 and C1 circuit is used to describe the battery diffusion phenomenon, and the R2 and C2 circuit is used to describe the battery polarization phenomenon. τ1 = P1C1, τ2 = R2C2.

[0033] The parameters of the second-order equivalent model of the battery are obtained:

[0034]

[0035] As a further improvement to the online lithium battery state-of-charge estimation method based on the state estimation method of the present invention:

[0036] Step S02 specifically involves:

[0037] S0201. Establishing an Unscented Kalman UKF Observer

[0038] The discrete state-space equation of the nonlinear system of a lithium battery can be simplified as follows:

[0039]

[0040] f is the nonlinear state equation of the system, g represents the nonlinear observation equation, and y n Let w be the voltage output observed at time n. n-1 and v n These are the system's process noise and the measurement noise during measurement, u n Let x be the current input value at time n. n The battery model state at time n:

[0041]

[0042] SOC n It is the SOC at time n. and The voltages of circuits R1C1 and R2C2 at time n are respectively.

[0043] 1) Algorithm Initialization

[0044]

[0045]

[0046] SOC0 is the initial state of charge obtained from the open-circuit voltage of the lithium battery, P0 is the state estimation variance, and Q is the process noise w. n The variance, R, is the measurement noise v. n variance

[0047] 2) Calculate the Sigma point set at time n-1:

[0048]

[0049] Among them, P n-1 Let be the variance of the state estimate at time n-1, where m is the dimension of the system state variables, and λ is a scaling factor defined as follows:

[0050] λ=α 2 (m+L)-m (53)

[0051] Let α = 0.1, L = 0.1m;

[0052] 3) Time update

[0053] Based on the above nonlinear equation (51), update all Sigma point sets backward:

[0054] x i,n|n-1 =f(x) i,n-1 ,u n-1 )+w n-1 ,i=1,…,2m (54)

[0055] Prior estimates of system state for

[0056]

[0057] in:

[0058]

[0059]

[0060] System prior estimate variance Represented as:

[0061]

[0062] Among them, Q n It is the covariance of the system process noise, and γ is the non-negative weighting coefficient;

[0063] 4) The corresponding measured value y i,n|n-1 Updated to:

[0064] y i,n|n-1 =g(x i,n|n-1 ,u n )+v n (59)

[0065] The estimated predicted value is obtained by using unscented changes; the average value of the terminal voltage is used. Covariance Represented as

[0066]

[0067] The joint covariance of the system's state and output variables at time n for:

[0068]

[0069] 5) Calculate the Kalman gain K n And correct the system status:

[0070]

[0071] System state estimation x n Revised to:

[0072]

[0073] System state covariance P n The correction is expressed as

[0074]

[0075] S0202, Introduction of Gated Recurrent Unit (GRU) Network

[0076] Define the input vector as

[0077] X = [X0, X1, ..., X 2m ] T (66)

[0078] After processing by the gated recurrent unit (GRU) network, the process noise is:

[0079]

[0080] G q The GRU network processing procedure representing process noise, X q This is the value of the process noise, h I,q h is the input value of the hidden layer unit in the GRU processing. o,q Y is the output value of the hidden layer unit. q X is the output of process noise. i,q X represents the current process noise state of the system (posterior probability, etc.). i W is the parameter matrix that needs to be learned, serving as the input for the next time step.

[0081] The measured noise is

[0082]

[0083] G r The GRU network processing procedure representing the measurement noise, X r It is the value of the measured noise, h I,r h is the input value of the hidden layer unit in the GRU processing. o,r Y is the output value of the hidden layer unit. r To measure the noise output, X i,r X represents the current measurement noise state of the system (posterior probability, etc.). i W is the input for the next time step. r It is the parameter matrix that needs to be learned;

[0084] The input feature X only requires one iteration to obtain the output vector and hidden vector; define r and q as the learning decay factors, and use... and Replace Q in the unscented Kalman UKF observer respectively n and R n :

[0085]

[0086]

[0087] As a further improvement to the lithium battery state-of-charge estimation method based on the state estimation method of the present invention:

[0088] Step S03 specifically involves:

[0089] The optimal parameter vector solution at time n, identified online by the adaptive attenuation factor recursive least squares (AFRLS) method in step S0103, is used. Then, the parameters R0, R1, R2, C1, and C2 of the second-order equivalent model of the battery are obtained by solving for them. These parameters are then substituted into the second-order equivalent model of the battery for updating, and the prior estimate of the system state is predicted. And estimate the variance P of the state at the previous time step. n-1 The Kalman gain K is calculated based on this. n The covariance matrix P n The system state is corrected by combining the difference between the actual battery terminal voltage and the predicted model terminal voltage.

[0090]

[0091] Obtain the state of charge (SOC) value at the next time step, output the SOC value, and feed it back to the adaptive decay factor recursive least squares (AFRLS) identification model for parameter identification, and update the unscented Kalman state observer (RGUKF) based on gated network optimization.

[0092] The beneficial effects of this invention are mainly reflected in:

[0093] The adaptive decay factor recursive least squares (AFRLS) method and the unscented Kalman state observer RGUKF based on gated network optimization used in this invention for joint prediction of SOC are more adaptable to various operating conditions, with lower computational complexity and higher accuracy compared to existing technologies. Under 25℃ conditions, the accuracy of this lithium battery state of charge estimation method is significantly improved compared to both the traditional extended Kalman (EKF) model and the unscented Kalman (UKF) model in constant rate discharge experiments, pulse discharge experiments, extreme initial error experiments, US06 experiments, FUDS experiments, and energy storage operating condition experiments. Attached Figure Description

[0094] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0095] Figure 1 This is a schematic diagram of the online lithium battery state of charge estimation process based on state estimation according to the present invention.

[0096] Figure 2 This is a schematic diagram of the equivalent circuit model of a second-order RC battery;

[0097] Figure 3 A framework for identifying parameters using the adaptive attenuation factor recursive least squares (AFRLS) method;

[0098] Figure 4 This invention describes the joint prediction process of the adaptive decay factor recursive least squares (AFRLS) and RGUKF methods.

[0099] Figure 5 This is a schematic diagram of the four input conditions in Experiment 1;

[0100] Figure 6 This is a schematic diagram of the online identification results of the adaptive attenuation factor recursive least squares (AFRLS) under the FUDS condition in Experiment 1.

[0101] Figure 7 The diagram shows the verification results (left figure) and errors (right figure) of the adaptive attenuation factor recursive least squares (AFRLS) method under the FUDS condition in Experiment 1.

[0102] Figure 8 This is a schematic diagram of the online identification results of the adaptive attenuation factor recursive least squares (AFRLS) under the DST condition in Experiment 1.

[0103] Figure 9 The diagram shows the identification results (left figure) and errors (right figure) of the adaptive attenuation factor recursive least squares method (AFRLS) under the DST condition in Experiment 1.

[0104] Figure 10 The diagram shows the verification results (left) and errors (right) of FUDS based on FRLS in Experiment 1.

[0105] Figure 11 This is a schematic diagram of the verification results (left figure) and errors (right figure) of the DST condition based on FRLS in Experiment 1;

[0106] Figure 12 This is a schematic diagram showing the SOC estimation results (left figure) and errors (right figure) under the DST condition in Experiment 1;

[0107] Figure 13This is a schematic diagram showing the SOC estimation results (left figure) and errors (right figure) of the FUDS condition experiment in Experiment 1;

[0108] Figure 14 This is a schematic diagram showing the SOC estimation results (left figure) and errors (right figure) of the US06 working condition experiment in Experiment 1;

[0109] Figure 15 This is a schematic diagram of the SOC estimation results (left figure) and errors (right figure) in the real energy storage scenario of Experiment 1. Detailed Implementation

[0110] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0111] Example 1

[0112] This invention provides an online lithium battery state-of-charge estimation method based on state estimation, such as... Figure 1 As shown, the OCV-SOC characteristic curves are obtained using a second-order RC equivalent circuit model, a small-current discharge method, and a hybrid pulse power performance testing method. Then, an optimized OCV-SOC curve is obtained using a differential evolution algorithm. Parameters are identified online using the Adaptive Attenuation Factor Recursive Least Squares (AFRLS) method, and the parameters of the unscented Kalman filter observer are updated. A time-series model gated cyclic unit network (GRU) is introduced to adaptively noise downtime, enabling online SOC prediction of the lithium battery as the output. Finally, the predicted SOC is combined with the OCV-SOC curve to predict the terminal voltage, which is then compared with the measured terminal voltage to form feedback for updating. The process includes the following steps:

[0113] Step 1: Obtain the characteristic curve L1 of discrete OCV-SOC and the characteristic curve L2 of continuous OCV-SOC.

[0114] The lithium battery charging and discharging platform was used to perform offline charging and discharging of the lithium battery, and the battery management system (BMS) was used to collect measured voltage, current and temperature data of the lithium battery under different operating conditions.

[0115] a. Obtain the characteristic curve L1 of discrete OCV-SOC using the static interval method.

[0116] 1) The interval resting method, also known as the Hybrid Pulse Power Characterization Test (HPPC), is used to charge the battery with constant current and constant voltage until the SOC reaches 100%, and then let it rest for more than 4 hours. The purpose of resting is to eliminate polarization effect and obtain accurate OCV value.

[0117] 2) Discharge using a constant current of 0.5C. Stop discharging when the discharged charge reaches 10% SOC, let it stand for more than 4 hours, and repeat the above discharge operation until the discharged charge reaches 100% SOC.

[0118] 3) After the discharge is complete, charge at a constant current of 0.5C. Stop charging when the charge reaches 10% SOC and let it stand for more than 4 hours. Repeat the above charging process until the charge reaches 100% SOC.

[0119] The entire process yields 22 OCV sampling points under different SOCs. After averaging the OCV sampling points under the same SOC, linear interpolation is used to obtain the characteristic curve L1 of the SOC-OCV under the interval static method.

[0120] b. Using the low-current discharge method, the characteristic curve L2 of continuous OCV-SOC is obtained by constant current discharge at a low current of 0.01C (C represents the battery capacity). The principle of the low-current discharge method is that the polarization effect generated inside the battery is minimal during the constant current discharge process of 0.01C. It is assumed that there is no polarization effect. Therefore, the measured value of the voltage during the constant current discharge process can be approximately equal to OCV. The characteristic curve L2 of SOC-OCV is obtained by collecting the voltage change data of the battery from full charge to full discharge.

[0121] Step 2: Based on the differential evolution (DE) method, the characteristic curve L1 of SOC-OCV obtained by the interval static method is combined with the continuity of the characteristic curve L2 of continuous OCV-SOC obtained by small current constant current discharge, and the optimized continuous OCV-SOC characteristic curve L3 is output.

[0122] Differential Evolution (DE) is an optimization algorithm based on swarm intelligence theory. It achieves global optimization of the objective function through operations such as mutation, crossover, and selection. The algorithm's principle is to perturb the direction of individual nodes to decrease their function values. Using characteristic curve L2 as the standard curve and discrete points on characteristic curve L1 as standard points, the algorithm moves the standard curve within a specified range to minimize the error between corresponding points on the standard curve and the standard points. Specifically, this includes:

[0123] Step 2.1: Establish the standard curve offset equation. Define the SOC variable as S and the OCV variable as O for the optimized continuous OCV-SOC characteristic curve L3. Then, the optimized SOC and OCV variables can be expressed as follows:

[0124] S = s s +h1·λ (1)

[0125] O = o s+h2·λ (2)

[0126] Among them, s s This represents the SOC variable value of the characteristic curve L2 obtained by constant current discharge with a small current of 0.01C, representing the continuous OCV-SOC characteristic curve. s The OCV variable values ​​h1·λ and h2·λ of the characteristic curve L2 obtained by constant current discharge with a small current of 0.01C represent two sets of offsets. The offset is controlled by setting a constant λ, which is 1.5. Considering the intuitive deviation estimation between the obtained standard curve and the standard point, h1∈[-0.1,0.1] and h2∈[-0.1,0.1] are defined.

[0127] Step 2.2, Define the objective function:

[0128]

[0129] The objective function represents minimizing the cumulative sum of the mean square error of the standard points obtained from the HPPC test by the optimized curve's SOC and OCV variables, where s h (i) represents the i-th standard SOC variable value of characteristic curve L2, o h (i) represents the i-th standard OCV variable value of characteristic curve L2, S(i) represents the SOC value of the i-th point of characteristic curve L3, and O(i) represents the OCV value of the i-th point of characteristic curve L3.

[0130] Step 2.3: Use the Differential Evolution (DE) algorithm to find the optimal solution in the state space for the objective function in Step 2.2 and the boundary conditions h1 and h2 described in Step 2.1.

[0131] Specifically, the minimum optimization problem with boundary conditions (DE):

[0132]

[0133] Where x1, x2, x3, ... x D Let S(i) and O(i) represent the inputs to the objective function within the state space defined by the boundary conditions. In this paper, these correspond to S(i) and O(i), respectively, where D represents the number of inputs. Let x represent the lower bound of the j-th input. j This represents the j-th input. The upper bound of the j-th input is given; the algorithm flow is as follows: Figure 2 As shown:

[0134] 1) Population Individual Initialization

[0135] Differential Evolution (DE) algorithm randomly generates multiple populations using multiple parameter vectors of dimension D. Generally, to cover more possibilities, the initial populations need to conform to a uniform probability distribution.

[0136]

[0137] Where, x j,i (0) represents the j-th gene on the i-th chromosome in generation 0. The total number of i is usually represented by NP, which represents the population size. rand(0,1) represents a random number uniformly distributed in the interval (0,1), ensuring that the population initialization conforms to a uniform probability distribution. These are the upper and lower bounds of the j-th gene on the i-th chromosome, respectively.

[0138] 2) Determine whether the termination condition is met.

[0139] The termination condition is the number of times the optimal solution of the objective function is found within the spatial range specified by the boundary conditions (according to empirical values, usually the number of times ≥ 200). If it is, the optimal solutions h1 and h2 of the offset are output, and then the characteristic curve L3 is output according to the optimization method in step 2.1 based on the characteristic curve L2. If not, proceed to step 3.

[0140] 3) Mutation operations and boundary condition checks

[0141] After initializing the population, mutation operations can be performed. The mutation strategy of the Differential Evolution (DE) algorithm is to perform a mutation by randomly subtracting individuals from two populations, and then using the resulting individuals, scaled according to the mutation exponent, as the mutation source and generating pseudo-offspring with the individuals to be mutated.

[0142] V i (k+1)=x r1 (k)+F·(x r2 (k)-x r3 (k)), i≠r1≠r2≠r3 (6)

[0143] Where, x r1 (k) represents the r1-th individual in the k-th generation of the population, x r2 (k) represents the r2th individual in the kth generation of the population, x r3 (k) represents the r3rd individual in the kth generation population, V i (k+1) represents pseudo-offspring individuals, and F is the scaling factor of the mutation source;

[0144] In the initial stage of the algorithm, to ensure the diversity of individuals in the population, an adaptive mutation scaling factor α is introduced as shown in the following formula, where F0 is the mutation operator, and G... U G represents the maximum number of generations, and G represents the current number of generations.

[0145]

[0146] F = F0·2 α (8)

[0147] Where F is the scaling factor of the mutation source;

[0148] The adaptive mutation factor ranges from F0 to 2F0. In the early stages, there are fewer generations and a larger mutation rate, ensuring the diversity of individuals. In the later stages, there are more generations and a smaller mutation rate, which can effectively preserve individuals that conform to the fitness function. In order to ensure the effectiveness of sample mutation, each new individual generated by mutation must be checked for boundary conditions. Only pseudo-offspring individuals generated by mutation that meet the boundary conditions are retained.

[0149] 4) Cross operations

[0150] To introduce interference vectors to increase diversity, the Kth generation population x... i (k) and the pseudo-offspring individuals V generated by the mutation i (k+1) Perform crossover operations between individuals:

[0151]

[0152] Among them, U j,i (k+1) represents the j-th allele on the i-th chromosome of the K+1 generation population after the crossover operation. CR∈[0,1] represents the crossover probability. The value of CR determines the frequency of use of the crossover operator. If its value is too small, it is not conducive to the development of population diversity. If its value is too large, it is easy to cause the excellent individuals in the population that meet the fitness function to be unable to be retained, which is not conducive to global optimization. Based on experience, the value is set to 0.3. In order to ensure the effectiveness of the crossover operation, that is, each crossover operation has at least one gene from the pseudo offspring generated by the mutation operation, the preset condition j=j is added. rand j rand Represents a random integer within the interval [0, D].

[0153] 5) Select Operation

[0154] The differential evolution (DE) algorithm's selection operation is centered on the objective function described in step 2.2. Only individuals that can reduce the mean squared error can enter the next generation of the population as offspring individuals.

[0155]

[0156] Then return to step 2) and continue to determine whether the termination condition is met;

[0157] The differential evolution (DE) algorithm finally outputs offsets h1 and h2 of -0.001654 and 0.00598, respectively. Based on the feature curve L2, the optimized continuous OCV-SOC feature curve L3 is obtained by outputting according to step 2.1.

[0158] Step 3: Online parameter identification using the adaptive attenuation factor recursive least squares method.

[0159] Step 3.1: Establish a single-input single-output model

[0160] The equivalent model of a lithium battery can be viewed as a single-input, single-output system with current as input and terminal voltage as output. The transfer function of this equivalent model can be expressed as:

[0161] F(n)=G(n)X(n) (11)

[0162] Where F(n) is the terminal voltage signal of the lithium battery equivalent model, X(n) represents the current signal input to the lithium battery equivalent model, F(n) and X(n) are both measurable discrete point values, and G(n) is the system transfer function.

[0163] The discrete transfer function of this lithium battery equivalent model can be expressed as:

[0164]

[0165] Where n is the number of zeros in the system, m is the number of poles in the system, and b1, b2, ... b m Let a1, a2, ... a be the coefficients corresponding to the zeros. n All are the coefficients of the corresponding poles, and Z is the complex variable in the Z-transform, which is re jΩ r is the real part, and j is the imaginary part. Ω represents the phase angle.

[0166] The system function G(Z) -1 The corresponding difference equation can be expressed as:

[0167]

[0168] Where, q(n)=[-f(n-1),-f(n-2),…,-f(nm),x(n-1),x(n-2),…,x(nm)] (14)

[0169] β T =[a1,a2,…,a m ,b1,b2,…,b m ] T (15)

[0170] Where x(n) is the current value at time n.

[0171] The actual measured value of the terminal voltage f(n) and the system-calculated q(n)β T There is a fitting error, denoted as o(n), i.e.

[0172] f(n)=q(n)β T +o(n) (16)

[0173] Therefore, the problem of equivalent circuit parameter identification can be transformed into finding the optimal parameter β that minimizes the sum of squared fitting errors J, i.e.:

[0174]

[0175] Wherein, minJ is the minimum sum of squared fitting errors.

[0176] Step 3.2: Multidimensional Model Based on Least Squares Method

[0177] Extend the input and output of this lithium-ion equivalent model to multidimensional variables:

[0178] F n =[f(1),f(2),…,f(n)] T =Q n β+O n (18)

[0179] Among them, O n Fitting error in a multidimensional function:

[0180] O n =[o(1),o(2),…,o(3)] (19)

[0181] β=[a1,a2,…,a m ,b1,b2,…,b m (20)

[0182]

[0183] Multidimensional variable evaluation function:

[0184]

[0185] use Let represent the optimal parameter vector solution. The least squares sum of the differences between the measured and estimated values ​​is obtained by searching the number space using the least squares method.

[0186] minJ=(F n -Q n β) T (F n -Q n β) (23)

[0187] make

[0188]

[0189] Simplify to get

[0190]

[0191] when When the term of office is completed,

[0192]

[0193] To find the derivative.

[0194] Step 3.3: Introduction of Recursive Least Squares (RLS)

[0195] The least squares method involves a huge amount of computation, and its real-time computational complexity increases dramatically in changing systems. Therefore, we introduce recursive least squares (RLS), assuming the optimal parameter vector solution at time n+1 is... At time n+1, q is q n+1 As can be seen from the idea of ​​Recursive Least Squares (RLS):

[0196]

[0197] Let be the optimal parameter vector solution at time n. can be The correction yields η(n+1), which is the correction amount at time n+1.

[0198] and:

[0199] Based on the principle of recursion

[0200]

[0201] F n+1 =[F n ,f n+1 ] T (30)

[0202] Next, define:

[0203]

[0204] so Derived as

[0205]

[0206] Therefore

[0207]

[0208] Ultimately obtainable

[0209]

[0210] definition

[0211] D n+1 =T n+1 q n+1 (35)

[0212] Therefore, it is simplified to

[0213]

[0214] For T n+1 Perform inverse expansion

[0215]

[0216] Bring it into D n+1

[0217]

[0218] Therefore

[0219]

[0220] Where E is and Identity matrices of the same order;

[0221] Therefore, the parameters of the Recursive Least Squares (RLS) method can be obtained as follows:

[0222]

[0223] Step 3.4: Online identification of adaptive attenuation factor using recursive least squares method

[0224] For a continuously running system, Recursive Least Squares (RLS) will update parameters indefinitely, causing the accumulated error to increase continuously. To address this issue, an adaptive decay factor c is introduced. n To reduce cumulative error:

[0225]

[0226] Among them, c n f is the true value of the system observation at time n. i Compared with the observed estimates The squared error value, where M is the window size for selecting the calculation sequence. The recursive process of the recursive least squares method is as follows: Figure 3 As shown.

[0227] The second-order equivalent model of the battery is as follows: Figure 4As shown, the circuit containing the battery's internal resistance R0, electrochemical polarization capacitor C1, and electrochemical polarization internal resistance R1 describes the battery's diffusion phenomenon, while the circuit containing the concentration polarization capacitor R2 and concentration polarization internal resistance C2 describes the battery's polarization phenomenon. The parameters at the current time n... When performing identification, the identification error is calculated by taking the current error value as the basis for the identification process. The identified R0, R1, R2, C1, and C2 are substituted into the second-order RC model of the battery. The current current is taken as input, and the predicted voltage is output. The identification error is obtained by subtracting the measured value from the predicted voltage, and this error is used in subsequent identification processes. When identifying parameters at time n, the mean square error (MSE) value is calculated using the identification errors from previous times n-1, n-2, ..., nM. The value c is then obtained as (1 - this MSE value). n c n The smaller the value, the less reliable the previous identification results are. To ensure convergence, we need to make c... n ≥0.95.

[0228] Parameter vector solution in Adaptive attenuation factor recursive least squares (AFRLS) for:

[0229]

[0230] Step 3.5: Establish the mathematical form of the second-order equivalent model of the battery.

[0231] according to Figure 4 The second-order equivalent model of the battery yields the Laplace equation for the equivalent model of the lithium battery.

[0232]

[0233] Among them, U ocv Let R0 represent the open-circuit voltage of the battery, R1 and C1 represent the internal resistance of the battery, R2 and C2 represent the diffusion phenomenon of the battery, and R2 and C2 represent the polarization phenomenon of the battery. U(s) represents the measurable terminal voltage of the battery, I is the current value, s is the Laplace operator, and the time constants τ1 = R1C1, τ2 = R2C2. Eliminating the denominators on both sides of the equation, we get:

[0234] U ocv +(τ1+τ2)U ocv s+τ1τ2U ocv s 2 =Iτ1τ2R0s 2 +(R1τ2+R2τ1+R0(τ1+τ2))Is+U+τ1τ2Us 2 +(τ1+τ2)Us+(R1+R2+R0)I (44)

[0235] Define q=τ1τ2, w=τ1+τ2, e=R1+R2+R0, r=R1τ2+R2τ1+R0(τ1+τ2), simplify to get

[0236] qU ocv s 2 +wU ocv s+U ocv =qR0Is 2 +rIs+eI+qUs 2 +wUs+U (45)

[0237] Discretization processing,

[0238]

[0239]

[0240] On the other hand, assuming

[0241]

[0242] Based on the parameter identification process of the Adaptive Decay Factor Recursive Least Squares (AFRLS) method, substituting into equation (45) yields:

[0243] U ocv，n -U n =d1(U n-1 -U ocv，n-1 )+d2(U n-2 -U ocv，n-2 )+d3I n +d4I n-1 +d5I n-2 (46)

[0244] in,

[0245]

[0246]

[0247] According to the Adaptive Decay Factor Recursive Least Squares (AFRLS): Solving for d1, d2, d3, d4, and d5 yields the parameters of the second-order equivalent model of the battery.

[0248] Define d0 = Δt 2 +wΔt+q, has

[0249]

[0250] q=d0d2

[0251] w=-d0(d1+2d2) / Δt

[0252] e = d0(d3 + d4 + d5) / Δt 2

[0253] r = -d0(d4 + 2d5) / Δt

[0254] R0 = d5 / d2

[0255] The parameters of the second-order equivalent model of the battery are obtained:

[0256]

[0257] Step 4: Establish an unscented Kalman state observer (RGUKF) based on gated network optimization.

[0258] Step 4.1: Establishment of the unscented Kalman UKF observer for lithium batteries

[0259] The discrete state-space equation of the nonlinear system of a lithium battery can be simplified as follows:

[0260]

[0261] f is the nonlinear state equation of the system, g represents the nonlinear observation equation, and y n Let w be the voltage output observed at time n. n-1 and v n These are the system's process noise and the measurement noise during measurement, u n Let x be the current input value at time n. n The battery model state at time n

[0262]

[0263] SOC n It is the SOC at time n. and The voltages of circuits R1C1 and R2C2 at time n are respectively.

[0264] 1) Algorithm Initialization

[0265]

[0266]

[0267] SOC0 is the initial value of the state of charge obtained from the open-circuit voltage of the lithium battery, P0 is the variance of the state estimation, w0 is the initial value of the variance of the process noise, v0 is the initial value of the measurement noise, and Q is the process noise. n The variance, R, is the measurement noise v. n The variance;

[0268] 2) Calculate the Sigma point set at time n-1:

[0269]

[0270] Among them, P n-1 Let m be the variance of the state estimate at time n-1, where m is the dimension of the system state variables, and λ is a scaling factor defined as:

[0271] λ=α 2 (m+L)-m (53)

[0272] α is non-negative and is used to control the state distribution of sample points, which can reduce the influence of higher-order matrices. In this paper, α = 0.1 and L = 0.1m are chosen.

[0273] 3) Time update:

[0274] Based on the above nonlinear equation (51), update all Sigma point sets backward:

[0275] x i,n|n-1 =f(x) i,n-1 ,u n-1 )+w n-1 ,i=1,…,2m (54)

[0276] Among them, w n-1 The noise at time n-1 is the process noise.

[0277] Prior estimates of system state for:

[0278]

[0279] in:

[0280]

[0281]

[0282] System prior estimate variance Represented as:

[0283]

[0284] Among them, Q n It is the covariance of the system process noise. The non-negative weighting coefficient γ is used to reduce peak error and improve the accuracy of covariance.

[0285] 4) The corresponding measured value y i,n|n-1 Updated to

[0286] y i,n|n-1 =g(x i,n|n-1 ,u n)+v n (59)

[0287] The estimated predicted value is obtained by using unscented changes; the average value of the terminal voltage is used. Covariance Represented as

[0288]

[0289] The next step is to obtain the joint covariance of the system's state and output variables at time n. for:

[0290]

[0291] 5) Calculate the Kalman gain K n And correct the system status:

[0292]

[0293] System state estimation x n Revised to:

[0294]

[0295] System state covariance P n The correction is expressed as

[0296]

[0297] 6) Introduction of Gated Recurrent Unit (GRU) Network

[0298] This study utilizes the noise state information from the previous time step to effectively guide the state parameters of the next time step, achieving adaptive online SOC estimation. Therefore, a gated recurrent unit (GRU) network is introduced. Compared to the traditional long-short-term memory (LSTM) network, the GRU network has one less gate and significantly fewer parameters, yet it achieves equivalent functionality. Considering the computational power and time cost, the more efficient GRU network was chosen.

[0299]

[0300] x t It is the input, h t These are hidden layer units, where sigm is the sigmoid function, tanh is the hyperbolic tangent function, and z... t It's an update gate that determines whether to ignore the current input, r tIt's a reset gate, controlling the hidden layer unit h from the previous moment. t-1 W xz W xr These are the update gate and the reset gate, respectively, for input x. t The parameter matrix, W hz W hr These are the parameter matrices of the hidden layer units at the previous time step, representing the update gate and reset gate, respectively. To represent the hidden layer value without considering the reset gate, ⊙ represents element-wise multiplication, W hh Let be the parameter matrix. The parameter matrix needs to be learned and updated.

[0301] In most cases, the factors influencing noise and the causes of noise generation are different. Therefore, it is not possible to directly use all influencing factors as input features to predict both types of noise, as this not only leads to uncertain interference but also causes fluctuations in the predicted values. This invention utilizes the residual value of system state updates to predict different noise values ​​and implements an adaptive method.

[0302] Define the input vector as

[0303] X = [X0, X1, ..., X 2m ] T (66)

[0304] After processing by the gated recurrent unit (GRU) network, the process noise is:

[0305]

[0306] G q The GRU network processing procedure representing process noise, X q This is the value of the process noise, h I,q h is the input value of the hidden layer unit in the GRU processing. o,q Y is the output value of the hidden layer unit. q This is the output of process noise. X i,q X represents the current process noise state of the system (posterior probability, etc.). i W is the input for the next time step. q It is the parameter matrix that needs to be learned.

[0307] The measured noise is:

[0308]

[0309] G r The GRU network processing procedure representing the measurement noise, X r It is the value of the measured noise, h I,r h is the input value of the hidden layer unit in the GRU processing.o,r Y is the output value of the hidden layer unit. r To measure the noise output. X i,r X represents the current measurement noise state of the system (posterior probability, etc.). i W is the input for the next time step. r It is the parameter matrix that needs to be learned.

[0310] The input feature X only needs one iteration to obtain the output vector and hidden vector.

[0311] Define r and q as the decay factors of learning, and use and Replace Q in the unscented Kalman UKF observer respectively n and R n :

[0312]

[0313]

[0314] Step 5: Output Results and Adaptive Decay Factor Recursive Least Squares (AFRLS) and Gated Network-Based Optimized Unscented Kalman State Observer (RGUKF) Correction.

[0315] The optimal parameter vector solution at time n, identified online by the adaptive decay factor recursive least squares (AFRLS) method in step 3.5, is used. Then, the parameters R0, R1, R2, C1, and C2 of the second-order equivalent model of the battery are obtained by solving for them. These parameters are then substituted into the second-order equivalent model of the battery for updating, and the prior estimate of the system state is predicted. And estimate the variance P of the state at the previous time step. n-1 The Kalman gain K is calculated based on this. n The covariance matrix P n The system state is corrected by combining the difference between the actual battery terminal voltage and the predicted model terminal voltage.

[0316]

[0317] Obtain the state of charge (SOC) value at the next time step, output the SOC value, and feed it back to the adaptive decay factor recursive least squares (AFRLS) identification model for parameter identification, and update the unscented Kalman state observer (RGUKF) based on gated network optimization.

[0318] Step 6: Use online

[0319] When using the SOC prediction method of this paper for the first time in a battery, it is necessary to import the battery SOC-OCV curve obtained in step 2, obtain the initial value of the battery SOC by the open circuit voltage method, and obtain the initial value of the battery's second-order model parameters by fitting the rebound characteristic curve after pulse discharge.

[0320] The measured current obtained by the current sensor and the measured voltage obtained by the voltage sensor at each moment, as well as the open-circuit voltage OCV (the open-circuit voltage OCV of the battery in the static state when inputting the SOC output at the previous moment) obtained by the curve in step 2, are respectively input to the online parameter identification model of AFRLS (the second-order model parameter initial value is used when using it) and the unscented Kalman observer RGUKF based on gated network optimization.

[0321] The online parameter identification model based on AFRLS updates the battery second-order model parameters in step 3. The unscented Kalman observer RGUKF based on gated network optimization uses the input and the updated battery second-order model to perform state prediction in steps 4 and 5, obtaining the SOC prediction value at each time step as the output. The output SOC value will be used as the input for the next time step.

[0322] experiment:

[0323] The experimental hardware configuration was as follows: Intel(R) Core(TM) i3-6100 CPU@3.70GHz processor; 8GB of memory; Intel(R) HD Graphics 530 graphics card; Windows 10 Enterprise Edition operating system; and MATLAB R2019B software.

[0324] In MATLAB, a model is built according to the process described in Example 1, and the following four operating conditions are used as inputs to the model built in Example 1:

[0325] (1) Energy storage operating conditions: including the voltage and current of the battery packs in operation of the energy storage power station throughout the entire day;

[0326] (2) Dynamic stress test (DST): This is an intermittent discharge operation state that includes the battery voltage and current throughout the entire simulation process;

[0327] (3) Federal Urban Driving Schedule (FUDS): This is a simulation of the battery operation in urban road sections, including the battery voltage and current throughout the simulation process.

[0328] (4) Aggressive driving conditions (The US06 Supplemental Federal Test Procedure, abbreviated as US06): This is a simulated operating state under extreme conditions such as high-speed driving and violent acceleration and deceleration, including the battery voltage and current throughout the simulation process.

[0329] Four operating conditions input as follows Figure 5 As shown.

[0330] First, according to the characteristic curve optimization method described in steps 1 and 2 of Example 1, the discrete OCV-SOC characteristic curve L1 is obtained through the Hybrid Pulse Power Performance Testing (HPPC) method, the continuous OCV-SOC characteristic curve L2 is obtained through small-current constant-current discharge, and the optimized SOC-OCV characteristic curve L3 is obtained. Based on characteristic curve L3, the parameters in the battery model are identified online using the Adaptive Attenuation Factor Recursive Least Squares (AFRLS) method. Combined with the gated network-optimized unscented Kalman state observer (RGUKF), the prediction of the battery SOC is completed. Specifically, this includes:

[0331] 1) Online parameter identification of adaptive attenuation factor recursive least squares (AFRLS) under FUDS and DST conditions

[0332] The online identification results of the adaptive attenuation factor using recursive least squares (AFRLS) under FUDS conditions are as follows: Figure 6 As shown;

[0333] The verification results (left figure) and error (right figure) of the adaptive attenuation factor recursive least squares (AFRLS) method under FUDS conditions are as follows: Figure 7 As shown;

[0334] The online identification results of the adaptive attenuation factor using recursive least squares (AFRLS) under DST conditions are as follows: Figure 8 As shown;

[0335] The results (left figure) and errors (right figure) of the adaptive attenuation factor recursive least squares (AFRLS) identification under DST conditions are as follows: Figure 9 As shown;

[0336] During the initial and final stages of FUDS and DST operating conditions, the current amplitude is relatively small, resulting in significant fluctuations in the identified parameter values ​​relative to the mean. As the operating conditions continue, in the middle stage, despite the drastic changes in current, the online identified parameter values ​​do not exhibit large fluctuations and tend to stabilize, allowing for effective real-time estimation of the equivalent circuit parameters.

[0337] 2) Online parameter identification of recursive least squares FRLS under FUDS and DST conditions

[0338] The FUDS condition verification results (left figure) and errors (right figure) based on FRLS are as follows: Figure 10 As shown,

[0339] The verification results (left figure) and errors (right figure) of DST condition based on FRLS are as follows: Figure 11 As shown.

[0340] Based on the above experimental results, the following conclusions can be drawn:

[0341] The online parameter identification method based on the second-order RC equivalent circuit model effectively corrects the terminal voltage observations under different operating conditions, avoiding deviations caused by changes in the internal chemical properties of the battery due to long-term use.

[0342] The online parameter identification method based on the equivalent model of the recursive least squares (AFRLS) method, while maintaining the same complexity as the traditional FRLS online parameter identification method, effectively reduces the accumulated error by using the equivalent model of the recursive least squares (AFRLS) method, and the output voltage is closer to the actual terminal voltage value. The accuracy is 25% to 35% higher than that of the FRLS method.

[0343] 3) Predict SOC using the method of this invention, EKF, and UKF under DST conditions.

[0344] The SOC estimation results (left figure) and error (right figure) under pulsed discharge (DST) conditions are as follows: Figure 12 As shown, a large estimation deviation occurs in the early stage of the pulsed discharge (DST) condition. As the condition progresses, the RGUKF model proposed in this paper is more effective in pulling back the estimation deviation than the other two traditional methods (UKF model and EKF model), and the overall error is controlled within 2%.

[0345] 4) Predict SOC using the method of this invention, EKF, and UKF under US06 operating conditions.

[0346] US06 operating condition test SOC estimation results (left figure) and errors (right figure) are as follows: Figure 13 As shown, under the US06 operating condition, the SOC estimation error based on the RGUKF model is the smallest, remaining within 2%. Specifically, the UKF model's estimation error is similar to that of the RGUKF model in the high SOC range, but the RGUKF model performs better. In the low SOC range, the estimation errors of the EKF and UKF models increase, and the estimation errors of the three methods become nearly identical. In summary, this is because the RGUKF model can continuously correct the covariance of system noise based on the residual between the actual measured voltage and the estimated voltage, maintaining a high convergence speed, thus improving the estimation accuracy of the RGUKF model as the operating condition progresses.

[0347] 5) Predict SOC using the method of this invention, EKF, and UKF under FUDS conditions.

[0348] The SOC estimation results (left figure) and errors (right figure) of the FUDS test are as follows: Figure 14 As shown, under FUDS conditions, the SOC estimation error based on the RGUKF model is the smallest and remains within 1.5%. Specifically, in the high SOC range, the estimation error of the RGUKF model is slightly higher than that of the UKF model, indicating that the UKF model performs better. As the operating conditions progress, the estimation error of the UKF model continuously increases, while the estimation accuracy of the RGUKF model gradually improves. In the low SOC range, the estimation errors of the EKF and UKF models increase, and the estimation errors of the three methods become nearly equal. In summary, this is because the RGUKF model can continuously correct the covariance of system noise based on the residual between the actual measured voltage and the estimated voltage, maintaining a high convergence speed, thus improving the estimation accuracy of the RGUKF model as the operating conditions progress.

[0349] 6) Predict SOC using the method of this invention, EKF, and UKF under real energy storage conditions.

[0350] The SOC estimation results (left figure) and errors (right figure) of the real-world energy storage scenario operating conditions are as follows: Figure 15 As shown, in the initial stage of the operating condition, the SOC estimation error accumulates and increases. When the battery is charging, the estimation errors of the RGUKF model, EKF model, and UKF model all decrease to varying degrees. However, as the battery resumes power supply, the error begins to increase again. The overall estimation errors of the EKF and UKF models continue to accumulate and increase, while only the RGUKF model can maintain a high estimation accuracy, remaining within 1.5%. The trend of SOC estimation error change in real energy storage operating conditions differs from that of the US06 and FUDS operating conditions. In the latter two operating conditions, the error value can be effectively pulled back during the estimation process, making the estimated value closer to the true value in the later stages of the operating condition. This is because the US06 and FUDS operating conditions consist of multiple cycles of a short period of drastically changing operating conditions, and the voltage and current values ​​have a cyclic period. Regardless of whether it is the traditional EKF model, the UKF model, or the RGUKF model in this paper, all can reduce the estimation error to varying degrees in the later stages of these multi-cycle operating conditions. However, real-world energy storage scenarios are unpredictable and constantly changing, and traditional estimation methods inevitably accumulate errors over time. The RGUKF model proposed in this paper can effectively solve this problem.

[0351] Based on the above experimental results, the following conclusions can be drawn:

[0352] (1) The RGUKF model based on AFRLS identification parameters (i.e. the method of this invention) can accurately predict the current state of charge change under set operating conditions, existing common cycle operating conditions and real energy storage scenario operating conditions, and the average error of the overall estimation is kept within 2%.

[0353] (2) By continuously correcting the covariance of system noise by the residual between the actual measured voltage and the estimated voltage, the convergence speed of the estimation is improved and the cumulative error is effectively reduced.

[0354] Finally, it should be noted that the above examples are merely some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. An online lithium battery state-of-charge estimation method based on state estimation, which first obtains the characteristic curves of discrete OCV-SOC through a hybrid pulse power performance test method. The characteristic curves of continuous OCV-SOC were obtained by the small current discharge method. The optimized continuous OCV-SOC characteristic curve is obtained based on the differential evolution method. Its features are: Includes the following steps: S01, using the characteristic curve As a standard, the equivalent circuit model of a second-order RC battery is parameterized using the Adaptive Fault Factor Recursive Least Squares (AFRLS) method. Initially, the open-circuit voltage (OCV) of the battery when it is at rest is input. Subsequent inputs are based on the battery voltage and current at each moment, as well as the predicted battery SOC from the previous moment, according to the optimized continuous OCV-SOC characteristic curve. The calculated open-circuit voltage OCV is used to output the parameter identification results of the second-order RC battery equivalent circuit model, including the characteristic curves. Parametric Ohmic Resistance of the Second-Order RC Battery Equivalent Circuit Model and second-order RC circuits , and , ; The process of parameter identification for the second-order RC battery equivalent circuit model using the adaptive attenuation factor recursive least squares (AFRLS) method in step S01 is as follows: S0101. Update the solution using Recursive Least Squares (RLS). ; make The recursive relation is obtained as follows: (40) in, This is the optimal parameter vector solution at time n; = (14) (21) in, The nonlinear state equations of the system; S0102. Introducing an adaptive attenuation factor into Recursive Least Squares (RLS). ; (41) It is the true value of the system observation at time n. Compared with the observed estimates The squared error value, where M is the window size for selecting the calculation sequence; Parameter vector solution in Adaptive attenuation factor recursive least squares (AFRLS) for: (42) in, For and Identity matrices of the same order; S0103. The mathematical form of the second-order equivalent model of the battery established by the adaptive attenuation factor recursive least squares method (AFRLS) is given by: = [ ],definition ,have ， ， ， ， ， ， in, , ， ， , , , ， The internal resistance of the battery point. For electrochemically polarized capacitors, For electrochemical polarization internal resistance, For concentration polarization internal resistance, For concentration polarization capacitors, , The loop is used to describe the diffusion phenomenon in the battery. , The circuit is used to describe the polarization phenomenon of the battery. , ; The parameters of the second-order equivalent model of the battery are obtained: (47)； S02: Based on the unscented Kalman UKF, introduce the GRU forget gate to establish an unscented Kalman state observer RGUKF model optimized by the gated network. Combine the second-order RC battery equivalent circuit model in step S01, input the measured voltage and current data of the lithium battery under different working conditions collected in real time and the battery SOC predicted at the previous moment, and predict the SOC value at the current moment. S03: Calculate the SOC value at the current moment based on the optimized continuous OCV-SOC characteristic curve. The parameters of the adaptive attenuation factor recursive least squares (AFRLS) and the unscented Kalman state observer RGUKF based on gated network optimization are adjusted and the model is updated by converting the parameters to the corresponding open-circuit voltage (OCV) as the input for the next time step.

2. The online lithium battery state-of-charge estimation method based on state estimation according to claim 1, characterized in that: The RGUKF model of the unscented Kalman state observer based on gated network optimization in step S02 is as follows: S0201. Establishing an Unscented Kalman UKF Observer The discrete state-space equation of the nonlinear system of a lithium battery can be simplified as follows: (48) Represents the nonlinear observation equation, The voltage output observed at time n. and These are the system's process noise and the measurement noise during measurement, respectively. Let n be the current input value at time n. The battery model state at time n: (49) It is the SOC at time n. and They are respectively at time n circuit and The voltage of the circuit; 1) Algorithm initialization is in: E (50) (51) It is the initial state of charge value obtained from the open-circuit voltage of the lithium battery. Q is the state estimation variance, and Q is the process noise. The variance, R is the measurement noise. The variance; 2) Calculate the Sigma point set at time n-1: (52) in, Let m be the variance of the state estimate at time n-1, where m is the dimension of the system state variables. It is a scaling factor, defined as: (53) Pick =0.1, L = 0.1m; 3) Time update: Based on the discrete state-space equation (48) of the nonlinear system of the lithium battery mentioned above, update all Sigma point sets backward: (54) Prior estimates of system state for (55) in: (56) (57) System prior estimate variance Represented as: (58) in, It is the covariance of the system process noise. These are non-negative weighting coefficients; 4) Corresponding measured values Updated to: (59) The estimated predicted value is obtained by using unscented changes; the average value of the terminal voltage is used. Covariance Represented as (60) The joint covariance of the system's state and output variables at time n for: (61) 5) Calculate the Kalman gain And correct the system status: (62) System state estimation Revised to: (63) System state covariance The correction is expressed as (64) S0202, Introduction of Gated Recurrent Unit (GRU) Network Define the input vector as (66) After processing by the gated recurrent unit (GRU) network, the process noise is: (67) The GRU network processing procedure representing process noise. This is the value of process noise. These are the input values ​​of the hidden layer units during the GRU processing. For hidden layer unit output values, For process noise output, Represents the current process noise state of the system. For the input of the next moment, It is the parameter matrix that needs to be learned; The measured noise is (68) The GRU network processing procedure representing the measurement noise. It measures the noise value. These are the input values ​​of the hidden layer units during the GRU processing. For hidden layer unit output values, To measure the noise output, This represents the current noise measurement status of the system. For the input of the next moment, It is the parameter matrix that needs to be learned; Input features The output and hidden vectors are obtained in just one iteration; r and q are defined as the learning decay factors, and... and Replacing the unscented Kalman UKF observers respectively and : ,(69) (70)。 3. The online lithium battery state-of-charge estimation method based on state estimation according to claim 2, characterized in that: Step S03 specifically involves: The optimal parameter vector solution at time n, identified online by the adaptive attenuation factor recursive least squares (AFRLS) method in step S0103, is used. Then, the parameters of the second-order equivalent model of the battery are obtained by solving. , , , , By substituting the second-order equivalent model of the battery into the updated model, the prior estimate of the system state is predicted. And estimate the variance based on the state at the previous time step. The Kalman gain was calculated based on this. Covariance Matrix The system state is corrected by combining the difference between the actual battery terminal voltage and the predicted model terminal voltage. (71) Obtain the state of charge (SOC) value at the next time step, output the SOC value, and feed it back to the adaptive decay factor recursive least squares (AFRLS) identification model for parameter identification, and update the unscented Kalman state observer (RGUKF) based on gated network optimization.

Images