A battery parameter dimensionality-raising evolution identification method based on feature empowerment

Abstract

The present application relates to the technical field of battery parameter identification, and relates to a battery parameter dimensionality evolution identification method based on feature empowerment, comprising the following steps: S101. Pulse power characteristic test and response voltage collection; S102. Battery equivalent circuit modeling; S103. Battery dynamic characteristic mathematical modeling; S104. Population initialization; S105. Population main target fitness evaluation; S106. Feature engineering construction based on expert knowledge; S107. Auxiliary optimization target construction and fitness evaluation; S108. Excellent population screening based on a Pareto multi-objective optimization mechanism; S109. Genetic evolution; S110. Termination condition judgment; whether the fitness tends to converge is judged, if the fitness tends to converge, the algorithm is terminated, and step S111 is entered; if the fitness has not converged, step S105 is returned to; S111. Parameter extraction. The present application can break through the local optimal limit, realize higher-precision parameter identification, enhance the stability of the identification process, and improve the optimization efficiency.

Description

Technical Field

[0001] This invention relates to the field of battery parameter identification technology, and more specifically, to a feature-empowered method for dimensional evolutionary identification of battery parameters. Background Technology

[0002] As a highly efficient and clean energy storage device, the battery has become a core energy carrier in key areas such as electric vehicles, large-scale energy storage power stations, and consumer electronics. The battery management system (BMS) is the brain that ensures the safe, reliable, and efficient operation of the battery pack. All its advanced management functions, such as accurate estimation of State of Charge (SOC) and State of Health (SOH), equalization control, thermal management, and fault diagnosis, highly depend on precise perception of the battery's internal state. To mathematically describe the complex electrochemical behavior of batteries, academia and industry commonly use equivalent circuit models, such as the Rint model, Thevenin model, and PNGV model. These models simulate the battery's internal dynamic characteristics, such as ohmic polarization, electrochemical polarization, and concentration polarization, through combinations of electronic components such as resistors and capacitors. Therefore, accurately identifying the parameters of these equivalent circuit models is fundamental to constructing high-fidelity battery digital twins and implementing all upper-level battery management system algorithms.

[0003] Currently, parameter identification methods for battery equivalent circuit models can be mainly divided into two categories: frequency domain methods and time domain methods. Frequency domain methods are represented by electrochemical impedance spectroscopy (EIS). This method applies sinusoidal perturbations of different frequencies to the battery and measures its impedance response, thereby fitting the model parameters in the frequency domain. EIS provides rich information about the various electrochemical processes within the battery and offers high identification accuracy. Its main drawback is its reliance on expensive dedicated measurement equipment, making online, real-time parameter identification impossible and difficult to deploy in practical battery management system products.

[0004] Time-domain methods directly utilize voltage and current time-series data collected during normal battery charging and discharging for parameter identification, making them more practically applicable in engineering. These methods typically transform the parameter identification problem into an optimization problem: finding a set of model parameters that best approximates the measured voltage response. Based on their optimization strategies, time-domain methods can be further subdivided into traditional optimization algorithms and heuristic intelligent optimization algorithms. Traditional optimization algorithms, such as Kalman filtering or recursive least squares, offer high computational efficiency. However, due to the inherent strong nonlinearity of battery models and the strong coupling between parameters, these gradient- or linearly based algorithms are sensitive to initial values ​​and prone to getting trapped in local optima, leading to insufficient identification accuracy. In contrast, heuristic intelligent optimization algorithms such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO), with their powerful global search capabilities and "model-independent" characteristics, demonstrate significant accuracy advantages in handling complex nonlinear optimization problems like battery parameter identification. They iteratively evolve, searching for the optimal parameter combination for the objective function within a broad solution space.

[0005] However, both the costly frequency domain method and the more widely used time domain method share a common, long-neglected limitation in their fundamental approach: their core idea is to construct the parameter identification problem as a single-objective optimization problem. For frequency domain methods, the optimization objective is usually to minimize the impedance spectrum fitting error (such as root mean square error RMSE or mean absolute error MAE) between the measured electrochemical impedance and the model impedance. In contrast, time domain methods, regardless of whether they employ traditional optimization algorithms or advanced heuristic intelligent optimization algorithms, almost invariably take the overall fitting error between the model output voltage and the actual measured voltage waveform as the sole optimization objective. Some literature has also attempted to construct new optimization objectives by introducing additional measurement hardware, but this undoubtedly greatly increases the system cost.

[0006] This "one-size-fits-all" evaluation method, while seemingly intuitive, presents two major technical bottlenecks: First, it makes the solution space of the optimization objective extremely complex, riddled with numerous pseudo-optimal solution traps. A relatively low overall fitting error may mask significant deviations in key physical features (such as voltage instantaneous response and dynamic relaxation processes), causing the algorithm to easily converge to a set of physically incorrect parameter solutions, fundamentally limiting the ceiling of identification accuracy. Second, because the optimization process is guided only by a single error metric, the algorithm's search behavior is highly random and blind, making it highly sensitive to the initial population and stochastic processes. This results in poor consistency of results across multiple independent identification experiments, insufficient stability and reproducibility, severely restricting its application in industrial scenarios requiring high reliability.

[0007] Therefore, how to overcome the inherent limitations of single-objective optimization without increasing any additional measurement hardware costs, and provide richer and more robust guiding information for the parameter identification process, is a key technical problem that urgently needs to be solved in the field of battery management. It has significant academic research value and broad industrial application prospects. Summary of the Invention

[0008] This invention aims to overcome the problems of low identification accuracy and unstable results in existing technologies that use evolutionary optimization algorithms for battery equivalent circuit model parameter identification. These problems are often caused by the high dimensionality and strong coupling of the parameters to be identified, as well as the single optimization objective. The invention provides a feature-enhanced method for dimensionality-upgrading evolutionary identification of battery parameters. This method, without increasing any additional measurement costs, constructs an auxiliary optimization objective based on the battery's physical characteristics, thus upgrading the original single-objective optimization problem to a multi-objective optimization problem. This significantly improves the algorithm's global search capability, convergence efficiency, and final identification quality.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A feature-empowered method for dimensional evolutionary identification of battery parameters includes the following steps:

[0011] S101. Pulse power characteristic test and response voltage acquisition;

[0012] S102. Battery equivalent circuit modeling;

[0013] S103. Mathematical modeling of battery dynamic characteristics; by deriving the battery state-space equation, the voltage response of the battery under different estimated parameter sets is calculated;

[0014] S104. Population initialization, forming an initial parent population of size 2*N;

[0015] S105. Assessment of population primary objective fitness;

[0016] S106. Feature engineering construction based on expert knowledge;

[0017] S107. Construction of Assisted Optimization Objectives and Evaluation of Their Fitness;

[0018] S108. Based on the Pareto multi-objective optimization mechanism, a new parent population of size N is formed by screening for excellent populations.

[0019] S109. Genetic evolution, through pairing selection, crossover, and mutation, forms a completely new offspring population of size N;

[0020] S110. Termination condition judgment; Determine whether the fitness is converging. If it is converging, the algorithm terminates and proceeds to step S111; if it has not yet converged, return to step S105.

[0021] S111. Extract parameters.

[0022] Furthermore, in step S101, the method for pulse power characteristic testing and response voltage acquisition includes:

[0023] Under constant temperature conditions, a programmable, periodic "rest-charging pulse-rest-discharging pulse-rest" pulse current is input to the battery to observe the battery's voltage dynamic response and relaxation process. Throughout the test, the battery test system continuously and synchronously records two time series data, namely the port current i(t) and the port voltage v(t). After the test, the two time series data are saved.

[0024] Furthermore, in step S102, a second-order RC equivalent circuit model is used to characterize the electrical characteristics of the battery. The second-order RC equivalent circuit model consists of a SOC capacitor (C), an ohmic internal resistance (R0), and two parallel RC networks connected in series.

[0025] The SOC capacitor (C) represents the battery's equivalent capacity and is used to simulate the battery's core energy storage function, namely, ampere-hour capacity.

[0026] The ohmic internal resistance (R0) represents the purely resistive part of the battery, including the electrode material, electrolyte, separator, and contact resistance between the components.

[0027] The first RC network (R1, C1) is used to simulate the fast dynamic response of the battery. R1 is the electrochemical polarization resistance and C1 is the electrochemical polarization capacitance.

[0028] The second RC network (R2, C2) is used to simulate the slow dynamic response of the battery. R2 is the concentration polarization resistor and C2 is the concentration polarization capacitor.

[0029] Furthermore, in step S103, the method for mathematical modeling the dynamic characteristics of the battery is as follows:

[0030] Let v1 and v2 be the voltages across capacitors C1 and C2, respectively, and V1(0) and V2(0) be the initial values ​​of the voltages across capacitors C1 and C2, respectively. OCV (0) represents the initial open-circuit voltage of the battery. According to Kirchhoff's laws, the continuous-time state-space equation of the model can be obtained as follows:

[0031] (1)

[0032] (2)

[0033] (3)

[0034] (4)

[0035] Discretizing the above equations, assuming the sampling time is Δt, the discretized state-space equation at the k-th sampling time is:

[0036] (5)

[0037] (6)

[0038] (7)

[0039] (8)

[0040] Among them, V OCV Mathematical relationship F between SOC OCV-SOC The set of battery model parameters θ to be identified is obtained in advance through OCV-SOC characteristic test. It includes {R0, R1, C1, R2, C2, V1(0), V2(0), C}. If the parameter to be identified θ and the current i(t) are known, the voltage response curve under the parameter can be derived by formula (5)-(8).

[0041] Furthermore, in step S104, the population consists of 2*N individuals, each of which is a complete set of model parameter vectors to be identified, P = [R0, R1, C1, R2, C2, V1(0), V2(0), C];

[0042] The population initialization method is as follows: First, set reasonable search boundaries for each parameter, with an upper bound UB and a lower bound LB; then, for each parameter to be generated, within its logarithmic interval [log 10 (LB), log 10 Generate a uniformly distributed random number within [(UB)]; finally, map this random number back to the original parameter value through antilogarithmic operation.

[0043] Furthermore, in step S105, the method for evaluating the primary objective fitness of the population is as follows:

[0044] For any individual P in the population i Substitute its parameters into the mathematical model of battery dynamic characteristics established in step S103, and then input the collected real current sequence i meas (t) is used as the model input to calculate a simulated response voltage curve v. sim (t), by comparing the simulated voltage with the real voltage v recorded at the same time point.meas The difference in (t) can be used to judge the accuracy of the parameters; for an individual Pi, its fitness value J in the primary objective can be used to judge the accuracy of the parameters. main The formula for calculating (Pi) is:

[0045] (9)

[0046] Where k is the discrete time step index, M is the total number of measurement data points, and v meas [k] represents the measured port voltage at time k, v sim(i) [k] represents the individual P. i The simulated voltage at time k obtained from the parameter derivation, and the fitness value J main The smaller (Pi) is, the more outstanding the individual is.

[0047] Furthermore, step S106 includes the construction of two key features:

[0048] High-precision R0 locking: According to the battery equivalent circuit model, when the port current i(t) undergoes a step change, the port voltage v(t) will produce an almost synchronous instantaneous jump. This jump is mainly caused by the ohmic internal resistance R0. The accuracy of R0 is judged by the fitting accuracy of individual voltage transient variables. Individuals with high fitting accuracy of voltage transient variables usually have high R0 accuracy.

[0049] Screening of individuals with high RC network parameter identification accuracy: During charging, discharging, and resting, the port voltage exhibits a slow change with a nonlinear, approximately exponential curve. The rate of change mainly depends on the time constant of the RC network. This characteristic reflects the dynamic characteristics within the model. By comparing the rate of change of the simulated voltage with that of the real voltage, individuals with high RC network parameter identification accuracy are screened.

[0050] Furthermore, in step S107, the method for constructing the auxiliary optimization target and evaluating its fitness is as follows: based on the two key features constructed in S106, they are quantified into two independent auxiliary optimization targets, and corresponding fitness functions are defined.

[0051] The first auxiliary optimization objective is the instantaneous response fitting error J. R0 This objective is used to quantify the accuracy of an individual in fitting instantaneous voltage response characteristics; the method is as follows: calculate the entire measured current sequence i meas The set of transient event points of [k] is K. transient For K transient For each time point k, calculate the actual voltage jump Δv. meas [k] and individual P i The corresponding analog voltage jump Δv sim(i) [k];

[0052] (10)

[0053] (11)

[0054] Instantaneous response fitting error J R0 Defined as the sum of squares error between the simulated voltage jump and the actual voltage jump at all transient event points, its calculation process is shown in formula (10); J R0 The smaller the value, the more accurately the individual identifies R0;

[0055] (12)

[0056] The second auxiliary objective is the fitting error of the dynamic rate of change, which is used to quantify the accuracy of an individual in fitting the dynamic characteristics of voltage polarization / relaxation. First, it is necessary to calculate the true port voltage rate of change during charging, discharging, and resting. and the rate of change of analog voltage Its discrete calculation formula is:

[0057] (13)

[0058] (14)

[0059] Where k is excluding K transient For time points other than the specified time point, Δt[k] represents the time interval between time point k and time point k-1; the fitting error J of the dynamic rate of change. RC J is defined as the sum of squares error between the simulated voltage rate of change sequence and the true voltage rate of change sequence. RC The smaller the value, the more accurate the individual's identification of the dynamic characteristics of the RC network;

[0060] (15)

[0061] At this point, each individual P in the population... i All of them obtained a three-dimensional target vector through evaluation [J] main (P i ),J R0 (P i ), J RC (P i )).

[0062] Furthermore, in step S108, the method for selecting superior populations based on the Pareto multi-objective optimization mechanism is as follows:

[0063] First, the base population for screening excellent populations is determined. In the first iteration, the base population is the initial parent population of size 2*N formed in step S104. In the second and subsequent iterations, the base population is a mixed population of size 2*N formed by merging the current parent population of total number N formed in step S108 of the previous iteration with the offspring population of total number N formed in step S109 of the previous iteration.

[0064] Then, based on the three-dimensional target vector calculated by S107, [J] main J R0 J RC The algorithm determines Pareto dominance for individuals. For any two individuals A and B, if all of A's objective values ​​are not inferior to B's, and at least one of A's objective values ​​is strictly superior to B's, then A is said to dominate B. The algorithm first identifies all individuals in the population that are not dominated by any other individual and assigns them to the first Pareto front. Then, it repeats this process among the remaining individuals, identifying the non-dominated solution set and assigning it to the second Pareto front, and so on, until all individuals are assigned a rank according to the stratification results. The smaller the rank, the better the individual.

[0065] Finally, based on the dual selection criteria of "prioritizing individuals with lower rank" and "if the ranks are the same, prioritizing individuals with greater crowding distance", N individuals are selected from the mixed population to form a new parent population.

[0066] Furthermore, in step S109,

[0067] The pairing selection method is as follows: Repeat the following three steps N times to build a pairing pool of size N:

[0068] S10911. Randomly and repeatedly select two individuals from a parent population of size N;

[0069] S10912. Compare the merits of these two individuals. The evaluation criteria are: prioritize the individual with a higher non-dominance level; if the levels are the same, choose the individual with a greater crowding distance.

[0070] S10913. Place the winning individual into the pairing pool;

[0071] The crossover operation is performed as follows: for each pair of parent individuals in the pairing pool, P1 = [p 1,1 ,...,p 1,j ] and P2=[p 2,1 ,...,p 2,j The simulated binary crossover operation is as follows:

[0072] S10921. Generate a random number u between [0,1];

[0073] S10922. Based on a preset distribution index η c Calculate an expansion factor β;

[0074] (16)

[0075] S10923. For each parameter j in the parent vector, generate two new child parameters c. 1,j and c 2,j :

[0076] (17)

[0077] (18)

[0078] The mutation operation is performed using a polynomial mutation operator that is compatible with the simulated binary crossover operation. For each offspring C generated after crossover, a small mutation probability p is applied. m Decide whether to mutate it; if mutation is performed, then independently perform the following operations on each parameter cj in the individual:

[0079] S10931. Generate a random number r between [0, 1];

[0080] S10932. Based on a preset variation distribution index η m Calculate a disturbance quantity δ;

[0081] (19)

[0082] S10933. Generate the new parameter c after mutation. ’ j Among them, UB j and LB j These are the upper and lower bounds for the search of parameter j; if c ’ j If it exceeds the boundary, set it as the boundary value:

[0083] (20)

[0084] By performing the above genetic evolution, a completely new offspring population of size N is eventually generated.

[0085] Compared with the prior art, the beneficial effects of the present invention are:

[0086] (1) The feature-enabled battery parameter dimensional evolution identification method in this invention breaks through the limitation of local optima and achieves higher precision parameter identification.

[0087] Existing technologies typically use the overall fitting error of the voltage response waveform under charge / discharge current excitation as the sole fitness function, such as Euclidean distance or root mean square error. This "one-size-fits-all" evaluation method fails to differentiate the response characteristics of the battery under different dynamic processes, resulting in a solution space riddled with local optima traps. Algorithms are highly prone to convergence to these traps during the optimization process, yielding a set of parameter solutions that, while having a smaller overall fitting error, deviate significantly from the actual battery behavior in key physical characteristics. This leads to insufficient accuracy in identifying the battery's equivalent electrical parameters.

[0088] Through feature engineering, key features with clear physical meaning are extracted from complex voltage response curves and constructed as independent auxiliary optimization objectives. In this way, the invention decomposes a single, fuzzy waveform fitting objective into multiple weakly correlated feature fitting objectives with clear expert experience backgrounds. This allows the multi-objective evolutionary algorithm to retain "shortcut individuals" that are not dominant in overall error but excel in a specific key feature. These individuals provide valuable evolutionary direction for the entire population, effectively guiding it out of local optima traps and towards the global optimum, thereby fundamentally improving the final identification accuracy of all model parameters.

[0089] (2) The feature-empowered battery parameter dimensional evolution identification method in this invention enhances the stability of the identification process and improves the optimization efficiency.

[0090] Traditional single-objective evolutionary optimization algorithms, due to their inherent "model-independent" and stochastic evolutionary characteristics, are highly sensitive to the randomness of the initial population and the evolutionary process. In multiple independent identification experiments, the algorithm may fall into different local optimum traps, resulting in significantly different parameter identification results, exhibiting poor stability and reproducibility. This uncertainty greatly limits the reliability of parameter identification methods in industrial applications.

[0091] The multi-objective optimization framework employed in this invention provides expert-level robust guidance for the evolutionary process by introducing the aforementioned auxiliary objectives. The elite preservation strategy of the multi-objective algorithm collaboratively preserves individuals with advantages across different objective dimensions, forming a diverse and high-quality parent population. This mechanism makes the algorithm's search path more robust. Therefore, even with different random initial populations and random evolutionary processes, the method of this invention can converge to the same global optimum with a higher probability, thereby obtaining highly consistent identification results and significantly enhancing the stability and reproducibility of the parameter identification process. Attached Figure Description

[0092] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0093] Figure 1 This is a flowchart of the feature-empowered battery parameter dimensional evolution identification method in this invention.

[0094] Figure 2 This is a circuit topology diagram of the experimental platform and battery equivalent circuit model in this invention.

[0095] Figure 3 This is a voltage response curve of the battery equivalent circuit model in this invention under charge and discharge pulses.

[0096] Figure 4 This is a schematic diagram of the excellent population selection based on the Pareto multi-objective optimization mechanism in step S108 of the present invention. Detailed Implementation

[0097] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0098] Example:

[0099] like Figure 1 As shown, the feature-empowered battery parameter dimensional evolution identification method in this embodiment includes the following steps:

[0100] S101. Pulse Power Characteristic Test and Response Voltage Acquisition

[0101] like Figure 2 As shown, an experimental platform consisting of a power electronic constant current source converter, a constant temperature chamber, the lithium-ion battery under test, and measuring equipment was constructed. The entire experiment was conducted at a constant temperature to eliminate temperature interference. Then, the dynamic electrical characteristics of the battery were tested using pulse power characteristic testing. This process involves inputting programmable, periodic "rest-charging pulse-rest-discharging pulse-rest" pulse currents into the battery and observing the battery's voltage dynamic response and relaxation process. Throughout the test, the battery testing system continuously and synchronously recorded two time-series data points: port current i(t) and port voltage v(t).

[0102] After the test, the collected i(t) and v(t) data are saved. These voltage and current data, which contain rich dynamic characteristics, will serve as the sole input for model parameter identification and feature target calculation in subsequent steps of this invention.

[0103] S102. Battery Equivalent Circuit Modeling

[0104] The circuit topology of the battery equivalent circuit model is as follows: Figure 2 As shown, it consists of a SOC capacitor, an ohmic internal resistor, and two parallel RC networks connected in series.

[0105] Among them, SOC capacitor C represents the battery's equivalent capacity, which simulates the battery's core energy storage function, namely ampere-hour capacity.

[0106] The ohmic internal resistance R0 represents the purely resistive part of the battery, including the electrode materials, electrolyte, separator, and contact resistance between the components. It causes the instantaneous step change in voltage when current flows through it.

[0107] The first RC network, R1 and C1, is used to simulate the rapid dynamic response of the battery. R1 is the electrochemical polarization resistance, and C1 is the electrochemical polarization capacitance. This network has a short time constant and mainly describes the rapid transfer and accumulation of charge at the electrode / electrolyte interface.

[0108] The second RC network, R2 and C2, is used to simulate the slow dynamic response of the battery. R2 is the concentration polarization resistor, and C2 is the concentration polarization capacitor. This network has a relatively long time constant and mainly describes the slow diffusion of lithium ions within the electrode active material and the process of concentration gradient establishment / dissipation.

[0109] This invention comprehensively considers the simulation accuracy and computational complexity of the model, and selects the aforementioned second-order RC equivalent circuit model to characterize the battery's electrical characteristics. This model is currently recognized by both industry and academia as a classic model that performs excellently in describing the dynamic response of power batteries. It can simultaneously simulate fast dynamic processes inside the battery, such as electrochemical polarization, and slow dynamic processes, such as concentration polarization, thereby more accurately reproducing the battery's voltage response curve under charge and discharge pulses.

[0110] S103. Mathematical Modeling of Battery Dynamic Characteristics

[0111] This step derives the battery state-space equation to calculate the battery's voltage response under different sets of estimated parameters. Let v1 and v2 be the voltages across capacitors C1 and C2, respectively, and V1(0) and V2(0) be the initial values ​​of the voltages across capacitors C1 and C2, respectively. OCV (0) represents the initial open-circuit voltage of the battery, which can be measured before the experiment begins. According to Kirchhoff's laws, the continuous-time state-space equation of the model is:

[0112] (1)

[0113] (2)

[0114] (3)

[0115] (4)

[0116] In practical digital computation, the above equations need to be discretized. Assuming the sampling time is Δt, at the k-th sampling time, the discretized state-space equation is:

[0117] (5)

[0118] (6)

[0119] (7)

[0120] (8)

[0121] Among them, V OCV Mathematical relationship F between SOC OCV-SOC The parameters can be obtained in advance through OCV-SOC characteristic testing. The set of battery model parameters θ to be identified in this invention includes {R0, R1, C1, R2, C2, V1(0), V2(0), C}. If the parameters to be identified θ and the current i(t) are known, the voltage response curve under these parameters can be derived through formulas (5)-(8).

[0122] S104. Population Initialization

[0123] This step generates a diverse initial solution set, or population, for the subsequent evolutionary identification algorithm. The population consists of 2*N individuals. A larger value for N yields better results but also increases computational cost. Therefore, in practical applications, the value of N is chosen based on the specific circumstances. Each individual is a complete set of parameter vectors for the model to be identified, P = [R0, R1, C1, R2, C2, V1(0), V2(0), C]. Given that battery parameters, such as resistance and capacitance, often have physical values ​​spanning multiple orders of magnitude, this embodiment employs a method of uniform random sampling in a logarithmic coordinate system.

[0124] First, set reasonable search boundaries for each parameter: an upper bound UB and a lower bound LB. The search range for parameters, derived by Na Jie based on actual industrial conditions, is narrowed down when there is more information and expanded when there is less information. Then, for each parameter to be generated, within its logarithmic interval [log...]... 10 (LB),log 10 A uniformly distributed random number is generated within [(UB)]. Finally, this random number is mapped back to the original parameter value through an antilogarithmic operation, i.e., a power of 10.

[0125] In this embodiment, N is set to 100, and the search boundary of R0 is 10. -3 -10-1 The R1 search boundary is 10. -4 -10 0 The R2 search boundary is 10. -4 -10 0 The search boundary for C1 is 10. -3 -10 4 The C2 search boundary is 10. 1 -10 5 The search boundary for V1(0) is 10. -1.5 -10 1.5 The search boundary for V2(0) is 10. -1.5 -10 1.5 The search boundary for C is 10. 4 -10 6 .

[0126] This method ensures that the generated initial population is uniformly distributed across all orders of magnitude, significantly improving population diversity. This lays a solid foundation for the algorithm to effectively avoid getting trapped in local optima and to converge quickly and stably to the global optimum.

[0127] S105. Population Aspect-Oriented Fitness Assessment

[0128] This step aims to quantitatively evaluate the performance of each individual in the population on the original primary objective, providing a criterion for the survival of the fittest in evolutionary algorithms. The core of the evaluation is to measure the accuracy of the model, composed of individual parameters, in reproducing the dynamic response of a real battery.

[0129] The evaluation strategy employs a comparison of simulated and measured response voltages. For any individual P in the population... i The parameters are then substituted into the mathematical model of battery dynamic characteristics established in step S103. Then, the collected real current sequence i... meas (t) is used as the model input to calculate a simulated response voltage curve v. sim (t). By comparing this simulated voltage with the real voltage v recorded at the same time point. meas The difference in (t) can be used to judge the accuracy of the set of parameters. For individual P i Its fitness value J for the primary objective main (P i The calculation formula is:

[0130] (9)

[0131] Where k is the discrete time step index, M is the total number of measurement data points, and v meas [k] represents the measured port voltage at time k, v sim(i) [k] represents the individual P. iThe simulated voltage at time k is obtained from the parameter derivation. Fitness value J main (P i The smaller the value, the more outstanding the individual.

[0132] S106. Feature Engineering Construction Based on Expert Knowledge

[0133] While the primary objective fitness function defined in step S105 can measure the consistency between the model output and the real data from a macroscopic perspective, it suffers from inherent information compression loss as a single evaluation dimension. It compresses the complex dynamic characteristics of the entire time series into a single scalar value, resulting in insufficient discrimination of the evaluation results and causing serious omissions of potentially high-quality individuals during the offspring selection process.

[0134] This step establishes two key features: locking individuals with high-precision R0 and screening individuals with high accuracy in identifying RC network parameters.

[0135] First, such as Figure 3 As shown, according to the battery equivalent circuit model, when the port current i(t) undergoes a step change, the port voltage v(t) will experience a nearly synchronous instantaneous jump. This jump is mainly caused by the ohmic internal resistance R0. Therefore, individuals with high fitting accuracy of the voltage transient variable usually have high R0 accuracy. This allows for the rapid locking of a high-precision R0 in the early stages of evolution, greatly simplifying the subsequent identification of other parameters and significantly accelerating the overall convergence process.

[0136] Secondly, such as Figure 3 As shown, during charging, discharging, and resting periods, the port voltage v(t) exhibits a slow, non-linear, approximately exponential change, with its rate of change primarily depending on the time constant of the RC network. This characteristic reflects the dynamic properties within the model. By comparing the rates of change of the simulated voltage with those of the actual voltage, individuals with high RC network parameter identification accuracy can be selected with a relatively high probability.

[0137] S107. Construction of Assisted Optimization Objectives and Fitness Evaluation

[0138] Based on the two key features constructed using S106, this step quantifies them into two independent auxiliary optimization objectives and defines corresponding fitness functions. This allows the evaluation of each individual's merits to be derived from a single primary objective J. main Elevate to multiple objectives.

[0139] The first auxiliary objective is the instantaneous response fitting error, which quantifies the accuracy of an individual in fitting the instantaneous voltage response characteristics. The entire measured current sequence i is calculated. meas The set of transient event points of [k] is K. transient For K transient For each time point k, calculate the actual voltage jump Δv.meas [k] and individual P i The corresponding analog voltage jump Δv sim(i) [k].

[0140] (10)

[0141] (11)

[0142] Auxiliary target J R0 Defined as the sum of squares error between the simulated voltage jump and the actual voltage jump at all transient event points, its calculation process is shown in formula (10). R0 The smaller the value, the more accurate the individual's identification of R0.

[0143] (12)

[0144] The second auxiliary objective is the fitting error of the dynamic rate of change. This objective is used to quantify the accuracy of an individual in fitting the dynamic characteristics of voltage polarization / relaxation. First, it is necessary to calculate the true port voltage rate of change during charging, discharging, and resting. and the rate of change of analog voltage Its discrete calculation formula is:

[0145] (13)

[0146] (14)

[0147] Where k is excluding K transient For time points other than the specified time point, Δt[k] represents the time interval between time point k and time point k-1. Auxiliary target J RC Defined as the sum of squares error between the simulated voltage rate of change sequence and the true voltage rate of change sequence. J RC The smaller the value, the more accurate the individual's identification of the dynamic characteristics of the RC network.

[0148] (15)

[0149] At this point, each individual P in the population... i All of them obtained a three-dimensional target vector through evaluation [J] main (P i ),J R0 (P i ), J RC (P i )).

[0150] S108. Selection of excellent populations based on Pareto multi-objective optimization mechanism

[0151] This step aims to select superior individuals from the current population using a Pareto-optimal multi-objective selection mechanism, while simultaneously retaining individuals that perform well on auxiliary objectives in the next generation. This allows them to drive the optimization process out of local optima during evolution. A schematic diagram of this step is shown below. Figure 4 As shown.

[0152] First, the base population for screening the superior population is determined. In the first iteration, the base population is the initial parent population of size 2*N formed in step S104. In the second and subsequent iterations, the base population is a mixed population of size 2*N formed by merging the current parent population of total number N formed in step S108 of the previous iteration with the offspring population of total number N formed in step S109 of the previous iteration.

[0153] Then, based on the three-dimensional target vector calculated by S107, [J] main J R0 J RC The algorithm determines Pareto dominance for individuals. For any two individuals A and B, A is said to dominate B if all of A's objective values ​​are not inferior to B's, and at least one of A's objective values ​​is strictly superior to B's. The algorithm first identifies all individuals in the population that are not dominated by any other individual and assigns them to the first Pareto front. Then, this process is repeated for the remaining individuals, identifying the non-dominated solutions and assigning them to the second Pareto front, and so on, until all individuals are assigned a rank according to the stratification results. The lower the rank, the better the individual.

[0154] Subsequently, the crowding distance between individuals within each level is calculated. This is achieved by summing the normalized distances between each individual within the same level and its two immediate neighbors in each target dimension. Individuals with larger crowding distances are located in sparser regions, contributing more to maintaining a broad distribution of the solution set, and are therefore more worthy of preservation.

[0155] Finally, based on the dual selection criteria of "prioritizing individuals with lower rank" and "if the ranks are the same, prioritizing individuals with greater crowding distance", N individuals are selected from the mixed population to form a new parent population.

[0156] S109. Genetic Evolution

[0157] This step aims to inherit the superior genes from the parent generation, namely parameter combinations, and to explore new regions of the parameter space by introducing random perturbations. This prevents the algorithm from converging prematurely to local optima and drives the population as a whole to evolve towards the global Pareto optimal front. Genetic evolution operations mainly include three core steps: pairing selection, crossover, and mutation.

[0158] Pairing Selection: Before generating offspring, individuals need to be selected from a parent population of size N for pairing and reproduction. This invention employs a binary tournament selection strategy. This strategy repeats the following three steps N times to build a pairing pool of size N:

[0159] S10911. Randomly and repeatedly select two individuals from the parent population;

[0160] S10912. Compare the merits of these two individuals, and the evaluation criteria are exactly the same as in S108: prioritize the individual with a higher non-dominance level; if the levels are the same, choose the individual with a greater crowding distance.

[0161] S10913. Place the winning individual into the pairing pool.

[0162] Crossover: The crossover operation simulates the exchange of chromosomes during biological reproduction, aiming to combine the superior traits of two parent individuals to produce potentially better offspring. This invention employs a simulated binary crossover operator optimized for real numbers. For each pair of parent individuals in the pairing pool, P1 = [p 1,1 ,...,p 1,j ] and P2=[p 2,1 ,...,p 2,j The simulated binary crossover operation is as follows:

[0163] S10921. Generate a random number u between [0,1];

[0164] S10922. Based on a preset distribution index η c This index controls the distance between offspring and parents, and an expansion factor β is calculated.

[0165] (16)

[0166] S10923. For each parameter j in the parent vector, generate two new child parameters c. 1,j and c 2,j :

[0167] (17)

[0168] (18)

[0169] Mutation: Mutation simulates gene mutation by applying a small random perturbation to the parameters of offspring individuals, introducing new genes into the population. It is crucial for maintaining population diversity and escaping local optima. This invention employs a polynomial mutation operator compatible with the simulated binary crossover operation. For each offspring individual C generated after crossover, a small mutation probability p is applied. mDecide whether to mutate it. If mutation is performed, then for each parameter c in the individual... j Independent operation:

[0170] S10931. Generate a random number r between [0, 1];

[0171] S10932. Based on a preset variation distribution index η m Calculate a disturbance quantity δ;

[0172] (19)

[0173] S10933. Generate the new parameter c after mutation. ’ j Among them, UB j and LB j These are the upper and lower bounds for the search of parameter j. If c ’ j If it exceeds the boundary, set it as the boundary value:

[0174] (20)

[0175] By performing the above genetic evolution, a completely new offspring population of size N is eventually generated.

[0176] S110. Termination Condition Determination and Optimal Solution Output

[0177] This step monitors the evolutionary process and terminates the algorithm when specific conditions are met. The algorithm terminates when the fitness of the individual performing best on the primary objective converges, i.e., when the fitness difference over Q consecutive iterations is less than a set threshold ε. If convergence has not yet occurred, the algorithm returns to S105 to continue iterating.

[0178] In this embodiment, the threshold ε is set to 1e. -4 Q takes the value 5, and the process is repeated 112 times.

[0179] S111. Extract parameters

[0180] After the algorithm converges, the individual with the best performance on the main objective is selected, and its parameters become the final parameter identification result. Table 1 shows a comparison of the parameter identification results of the algorithm proposed in this patent and traditional evolutionary optimization algorithms. It is evident that the identification accuracy and stability are greatly improved.

[0181] Table 1

[0182]

[0183] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feature-based method for dimensional evolutionary identification of battery parameters, characterized in that, Includes the following steps: S101. Pulse power characteristic test and response voltage acquisition; S102. Battery equivalent circuit modeling; The electrical characteristics of the battery are characterized by a second-order RC equivalent circuit model, which consists of a SOC capacitor C, an ohmic internal resistance R0, and two parallel RC networks connected in series. S103. Mathematical modeling of battery dynamic characteristics; by deriving the battery state-space equation, the voltage response of the battery under different estimated parameter sets is calculated; S104. Population initialization, forming an initial parent population of size 2*N; S105. Assessment of population primary objective fitness; S106. Feature engineering construction based on expert knowledge; including the construction of two key features: high-precision R0 locking and screening of individuals with high RC network parameter identification accuracy; S107. Construction of Auxiliary Optimization Objectives and Evaluation of Fitness: Based on the two key features constructed in S106, they are quantified into two independent auxiliary optimization objectives, and corresponding fitness functions are defined; the first auxiliary optimization objective is the instantaneous response fitting error J. R0 The first objective is used to quantify the accuracy of an individual in fitting instantaneous voltage response characteristics; the second auxiliary objective is the fitting error of the dynamic rate of change, which is used to quantify the accuracy of an individual in fitting voltage polarization / relaxation dynamic characteristics. S108. Based on the Pareto multi-objective optimization mechanism, a new parent population of size N is formed by screening for excellent populations. S109. Genetic evolution, through pairing selection, crossover, and mutation, forms a completely new offspring population of size N; S110. Termination condition judgment; Determine whether the fitness is converging. If it is converging, the algorithm terminates and proceeds to step S111; if it has not yet converged, return to step S105. S111. Extract parameters.

2. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 1, characterized in that, In step S101, the method for pulse power characteristic testing and response voltage acquisition includes: Under constant temperature conditions, a programmable, periodic "rest-charge pulse-rest-discharge pulse-rest" pulse current is input to the battery to observe the battery's voltage dynamic response and relaxation process. Throughout the test, the battery test system continuously and synchronously records two time series data, namely the port current i(t) and the port voltage v(t). After the test, the two time series data are saved.

3. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 2, characterized in that, In step S102; The SOC capacitor C represents the battery's equivalent capacity, used to simulate the battery's core energy storage function, namely, ampere-hour capacity. The internal resistance R0 represents the purely resistive part of the battery, including the electrode material, electrolyte, separator, and contact resistance between the components. The first RC network (R1, C1) is used to simulate the fast dynamic response of the battery. R1 is the electrochemical polarization resistance and C1 is the electrochemical polarization capacitance. The second RC network (R2, C2) is used to simulate the slow dynamic response of the battery. R2 is the concentration polarization resistor and C2 is the concentration polarization capacitor.

4. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 3, characterized in that, In step S103, the method for mathematical modeling the dynamic characteristics of the battery is as follows: set up v 1 and v 2 are capacitors C 1 and C 2. Voltage across the terminals V 1(0) and V 2(0) represent capacitors respectively. C 1 and C 2. The initial value of the voltage across the terminals. V OCV (0) represents the initial open-circuit voltage of the battery. According to Kirchhoff's laws, the continuous-time state-space equation of the model can be obtained as follows: Discretizing the above equations, assuming the sampling time is Δt, the discretized state-space equation at the k-th sampling time is: Among them, V OCV Mathematical relationship F between SOC OCV-SOC The set of battery model parameters θ to be identified is obtained in advance through OCV-SOC characteristic test. It includes {R0, R1, C1, R2, C2, V1(0), V2(0), C}. If the parameter to be identified θ and the current i(t) are known, the voltage response curve under the parameter can be derived by formula (5)-(8).

5. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 4, characterized in that, In step S104, the population consists of 2* N It consists of individual entities, each of which is a complete set of parameter vectors for the model to be identified. P = [ R 0, R 1, C 1, R 2, C 2, V 1(0), V 2(0), C ]; The method for population initialization is as follows: First, set reasonable search boundaries for each parameter, including an upper bound. UB and the lower realm LB Then, for each parameter to be generated, in its logarithmic interval [log 10 ( LB ), log 10 ( UB Generate a uniformly distributed random number within the []; finally, map the random number back to the original parameter value through antilogarithmic operation.

6. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 5, characterized in that, In step S105, the method for evaluating the primary objective fitness of the population is as follows: For any individual P in the population i Substitute its parameters into the mathematical model of battery dynamic characteristics established in step S103, and then input the collected real current sequence i meas (t) is used as the model input to calculate a simulated response voltage curve v. sim (t), by comparing the simulated voltage with the real voltage v recorded at the same time point. meas The difference in (t) can be used to judge the accuracy of the parameter; for individual P i Its fitness value J in the primary objective main (P i The calculation formula is: Where k is the discrete time step index, M is the total number of measurement data points, and v meas [k] represents the measured port voltage at time k, v sim(i) [k] represents the individual P. i The simulated voltage at time k obtained from the parameter derivation, and the fitness value J main The smaller (Pi) is, the more outstanding the individual is.

7. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 6, characterized in that, In step S106: High-precision R0 locking: According to the battery equivalent circuit model, when the port current i(t) undergoes a step change, the port voltage v(t) will produce an almost synchronous instantaneous jump. This jump is mainly caused by the ohmic internal resistance R0. The accuracy of R0 is judged by the fitting accuracy of individual voltage transient variables. Individuals with high fitting accuracy of voltage transient variables have higher R0 accuracy. Screening of individuals with high RC network parameter identification accuracy: During charging, discharging, and resting, the port voltage exhibits a slow change with a nonlinear, approximately exponential curve. The rate of change mainly depends on the time constant of the RC network. This characteristic reflects the dynamic characteristics within the model. By comparing the rate of change of the simulated voltage with that of the real voltage, individuals with high RC network parameter identification accuracy are screened.

8. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 7, characterized in that, In step S107, Instantaneous response fitting error J R0 The calculation method is as follows: calculate the entire measured current sequence i meas The set of transient event points of [k] is K. transient For K transient For each time point k, calculate the actual voltage jump Δv. meas [k] and individual P i The corresponding analog voltage jump Δv sim(i) [k]; Instantaneous response fitting error J R0 Defined as the sum of squares error between the simulated voltage jump and the actual voltage jump at all transient event points, its calculation process is shown in formula (10); J R0 The smaller the value, the more accurately the individual identifies R0; The method for calculating the fitting error of the dynamic change rate is as follows: First, it is necessary to calculate the actual port voltage change rate during charging, discharging, and resting periods. and the rate of change of analog voltage ; Its discrete calculation formula is: Where k is excluding K transient For time points other than the specified time point, Δt[k] represents the time interval between time point k and time point k-1; the fitting error J of the dynamic rate of change. RC J is defined as the sum of squares error between the simulated voltage rate of change sequence and the true voltage rate of change sequence. RC The smaller the value, the more accurate the individual's identification of the dynamic characteristics of the RC network; At this point, every individual in the population... P i All of them yielded a three-dimensional target vector through evaluation. J main ( P i ), J R0 ( P i ), J RC ( P i )).

9. The battery parameter dimensionality-upgrading evolution identification method based on feature empowerment according to claim 8, characterized in that, In step S108, the method for selecting superior populations based on the Pareto multi-objective optimization mechanism is as follows: First, the base population for screening excellent populations is determined. In the first iteration, the base population is the initial parent population of size 2*N formed in step S104. In the second and subsequent iterations, the base population is a mixed population of size 2*N formed by merging the current parent population of total number N formed in step S108 of the previous iteration with the offspring population of total number N formed in step S109 of the previous iteration. Then, based on the three-dimensional target vector calculated by S107, [J] main J R0 J RC The algorithm assesses Pareto dominance among individuals. For any two individuals A and B, A is said to dominate B if all of A's objective values ​​are not inferior to B's, and at least one of A's objective values ​​is strictly superior to B's. The algorithm first identifies all individuals in the population that are not dominated by any other individual and assigns them to the first Pareto front. Then, this process is repeated among the remaining individuals to identify the non-dominated solution set and assign it to the second Pareto front, and so on, until all individuals are assigned a rank according to the stratification results. The lower the rank, the better the individual. Finally, based on the dual selection criteria of "prioritizing individuals with lower ranks" and "if ranks are the same, prioritizing individuals with larger crowding distances," N individuals are selected from the mixed population to form a new parent population.

10. The battery parameter upsizing and evolutionary identification method based on feature empowerment according to claim 9, characterized in that, In step S109, The pairing selection method is as follows: Repeat the following three steps N times to build a pairing pool of size N: S10911. Randomly and repeatedly select two individuals from a parent population of size N; S10912. Compare the merits of these two individuals. The evaluation criteria are: prioritize the individual with a higher non-dominance level; if the levels are the same, choose the individual with a greater crowding distance. S10913. Place the winning individual into the pairing pool; The crossover operation is performed as follows: for each pair of parent individuals in the pairing pool, P1 = [p 1,1 ,...,p 1,j ] and P2=[p 2,1 ,...,p 2,j The simulated binary crossover operation is as follows: S10921. Generate a random number u between [0,1]; S10922. Based on a preset distribution index η c Calculate an expansion factor β; S10923. For each parameter j in the parent vector, generate two new child parameters c. 1,j and c 2,j : The mutation operation is performed using a polynomial mutation operator that is compatible with the simulated binary crossover operation. Each offspring generated after crossover is mutated with a relatively small probability. p m Decide whether to mutate it; if mutation is performed, then independently perform the following operations on each parameter cj in the individual: S10931. Generate a random number r between [0,1]; S10932. Based on a preset variation distribution index... η m Calculate a disturbance quantity δ; S10933. Generate the new parameters after mutation. c ’ j Among them, UB j and LB j These are the upper and lower bounds for the search of parameter j; if c ’ j If it exceeds the boundary, set it as the boundary value: By performing the above genetic evolution, a completely new offspring population of size N is eventually generated.

Images