Battery fast charging strategy design method based on feature driving and machine learning closed loop optimization

Abstract

The application relates to a feature-driven and machine learning-based closed-loop optimization battery fast charging strategy design method, and belongs to the battery fast charging optimization field, and comprises the following steps: S1: based on COED, m charging protocols are randomly selected from all charging protocols, and a battery experiment device is used to perform n times of charging and discharging cycles on the battery; S2: a feature extractor is used to extract features related to the battery life corresponding to the m charging protocols from the previous n times of charging and discharging cycles as the output of a training set of a black box optimization model, and the charging protocols as the input jointly form the training set, and the black box optimization model is trained; S3: the trained black box optimization model is used to predict the features of all charging protocols, and the protocols to be tested in the next round are acquired; S4: the protocols to be tested in the next round and the features related to the battery life thereof are added to the original training set, the black box optimization model is retrained, and steps S3-S4 are repeated until a condition is met or an optimization target is reached.

Description

Technical Field

[0001] This invention belongs to the field of battery fast charging strategy optimization, and relates to a battery fast charging strategy design method based on aging characteristics, machine learning and closed-loop optimization. Background Technology

[0002] The vigorous development and popularization of electric vehicles is of great significance for mitigating the greenhouse effect and protecting the environment, which has also promoted the development and progress of lithium-ion batteries (LIBs) to a certain extent. Due to battery limitations, electric vehicles suffer from range anxiety and long charging times, while using high-energy-density batteries and fast-charging strategies can effectively solve these problems.

[0003] Commonly used charging protocols include constant current constant voltage (CC-CV) and constant power constant voltage (CP-CV), which consist of a constant current (CC) or constant power (CP) stage and a CV stage to avoid overcharging. Another charging protocol is called multi-stage constant current constant voltage charging (MSCC), which consists of multiple constant current charging stages and one constant voltage charging stage. It requires determining the optimal charging current for each stage of the battery. The switching time from one stage to another during charging is determined by the time to reach a preset charging voltage or state of charge (SOC), and the charging rate generally decreases progressively. Some studies have found that, in the case of 10-minute fast charging, if the charging rate is switched after reaching a certain SOC, the optimal charging protocol is obtained when the charging rates of each stage are similar. MSCC charging protocols have the following problems: high dimensionality of the charging protocol space, limited quantitative information on battery degradation mechanisms, and significant differences between batteries. Therefore, conducting a single lifetime assessment experiment requires months to years of extensive experimentation, necessitating optimization of the experimental process.

[0004] However, simultaneously optimizing multiple parameters or high-dimensional parameters is a challenging task that requires significant time for experimentation. Traditional solutions include model-based optimization methods and grid search methods, both of which have obvious drawbacks. Model-based optimization methods can utilize models to avoid experimentation, but model biases can lead to inaccurate optimization results. Grid search methods, on the other hand, require a large number of experiments, the number of which depends on the number of parameters and the preset range of each parameter. This necessitates a substantial investment of time and effort in optimization experiments. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a closed-loop optimization battery fast charging strategy design method based on feature-driven and machine learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A closed-loop optimization method for battery fast charging strategy design based on feature-driven and machine learning includes the following steps:

[0008] S1: Based on the closed-loop optimization experimental design method COED, m charging protocols are randomly selected from all charging protocols, and the battery experimental device is used to perform n charge-discharge cycles on the battery.

[0009] S2: During the cycle, the feature extractor extracts the battery life-related features corresponding to the m charging protocols from the first n charge-discharge cycles, and uses them as the output of the training set of the black-box optimization model. The output of the black-box optimization model is combined with the charging protocols used as input to form the training set, and the black-box optimization model is trained.

[0010] S3: Utilize the trained black-box optimization model to predict the features of all charging protocols and obtain the protocols to be tested in the next round.

[0011] S4: Add the protocol to be tested in the next round and its battery life-related features to the original training set to obtain a new training set to retrain the black-box optimization model. Repeat steps S3-S4 until the conditions are met or the optimization goal is achieved.

[0012] Furthermore, all the charging protocols constitute a charging protocol space, in which the charging protocols aim to charge the battery's state of charge (SOC) from 0% to 80% within 10 minutes through multiple stages of constant current charging. During this process, j different magnitudes of current are used to charge the battery in stages.

[0013] Furthermore, the charging protocol is Multi-Stage Constant Current Constant Voltage (MSCC), which consists of multiple constant current charging stages followed by a constant voltage charging stage.

[0014] When j=3, the charging protocol charges the battery's State of Charge (SOC) from 0% to 80% in 3 steps within 10 minutes. The parameters of the charging protocol are CC1′ and CC2′, which are also the two-dimensional inputs of the black-box optimization model. Specifically, the charging rate is CC1′ in the 0-26.67% SOC range, CC2′ in the 26.67%-53.34% SOC range, and CC3′ in the 53.34%-80% SOC range. The magnitude of CC3′ is determined by CC1′ and CC2′, and is expressed as:

[0015]

[0016] When j=4, the charging protocol charges the battery's State of Charge (SOC) from 0% to 80% in 4 steps within 10 minutes. The parameters of the charging protocol are CC1, CC2, and CC3, which are also the three-dimensional inputs of the black-box optimization model. Specifically, the charging rate is CC1 for the 0-20% SOC range, CC2 for the 21%-40% SOC range, CC3 for the 41%-60% SOC range, and CC4 for the 61%-80% SOC range. The value of CC4 is determined by CC1, CC2, and CC3, as shown in the following formula:

[0017]

[0018] When j=5, the charging protocol charges the battery's State of Charge (SOC) from 0% to 80% in 5 steps within 10 minutes. The parameters of the charging protocol are CC1″, CC2″, CC3″, and CC4″, which are also the four-dimensional inputs of the black-box optimization model. Specifically, the charging rate is CC1″ for the 0-16% SOC range, CC2″ for the 16%-32% SOC range, CC3″ for the 32%-48% SOC range, CC4″ for the 48%-64% SOC range, and CC5″ for the 64%-80% SOC range. The value of CC5″ is determined by CC1″, CC2″, CC3″, and CC4″, as shown in the following formula:

[0019]

[0020] Furthermore, the feature extractor extracts m features corresponding to charging protocols as variance features, and calculates the logarithm of the variance of the difference between the discharge capacity-voltage curves of the nth and pth charging cycles as Log(Var(Qn)). -p (V))) is used as a characteristic to characterize battery life, where p is a positive integer less than n.

[0021] Furthermore, the black-box optimization model is a Bayesian optimization model, which consists of a Gaussian process regression model (GPR) and a data acquisition function.

[0022] In step S2, the m three-dimensional charging protocols CC1, CC2, CC3 and their corresponding log-variance features Log(Var(Qn)) are... -p (V))) is used as the training set to train the GPR model, thereby outputting the log-variance features of all three-dimensional charging protocols;

[0023] The acquisition function is used to determine which protocols to test in the next round.

[0024] Furthermore, the GPR is used to determine the eigenvalues ​​corresponding to each protocol. The input is the three-dimensional protocol CC1 to CC3, and the output is the log-variance Log(Var(Q)).n-p (V)));

[0025] When applying GPR, the actual output y after considering noise is expressed as:

[0026] y i =f(x) i )+ε i

[0027] Where (x) i y i Let f represent the i-th data point, f(x) represent the feature map, and f(x) represent the i-th data point. i ) represents the output of the Gaussian model; ε i Indicates observation noise, and

[0028] The set of any finite inputs X = {x1, x2, ..., x...} n} d*n Both follow a joint Gaussian probability distribution p(f(x1), f(x2)). z The mean function and covariance function of f(x), ..., f(xn) are respectively:

[0029] m(x) = E[f(x)]

[0030] k(x,x′)=E[(f(x)-m(x))(f(x′)-m(x′))]

[0031] Since there is no prior information about the distribution of f(x), m(x) is set to 0, and the distribution of y can be written as:

[0032]

[0033] Where K ff Let be an N*N dimensional matrix, and [K ff ] i，j =k(x i x j );

[0034] For a new input Its joint distribution with the known observed target values ​​can be written as:

[0035]

[0036] in [K ff ] i，j =k(x i x j );

[0037] Conditional distribution p(f) * |X, y, X* It follows a normal distribution:

[0038]

[0039] in, Let cov(f) represent the mean of the posterior distribution. * The variance of the posterior distribution is represented by the following formula:

[0040]

[0041]

[0042] Furthermore, the kernel functions that can be used in the GPR model include, but are not limited to, the following:

[0043] The first kernel function is the Gaussian kernel RBF, given by the following formula:

[0044]

[0045] Where ||xx′|| 2 Let σ represent the square of the magnitude of vector xx′, and let σ represent the standard deviation.

[0046] The second kernel function is the Matern kernel, given by the following formula:

[0047]

[0048] It has positive parameters v and l, where Γ(v) is a Γ function, K ν It is a Bessel function of the second kind, and d(x, x′) represents the Euclidean distance;

[0049] The third kernel function is the white noise kernel function:

[0050]

[0051] Where, σ m is a constant representing the level of white noise.

[0052] Furthermore, the acquisition function is a confidence interval upper bound function UCB, an improved probability function PI, or an expected increment function EI, where:

[0053] The UCB acquisition function is as follows:

[0054] UCB(x) = μ(x) + βσ(x)

[0055] Where μ(x) and σ(x) are the mean and standard deviation of the posterior distribution, respectively, and β is a hyperparameter to be optimized to balance exploration and development. Exploration refers to exploring untested parameter spaces with high uncertainty, while development refers to testing promising regions based on completed experimental results. As the number of steps increases, the focus shifts from exploration to development, and β is gradually reduced, resulting in:

[0056] UCB(x)=μ(x)+β0ε k σ(x)

[0057] Where β0, ε k All are hyperparameters, k represents the number of optimizations, and ε takes values ​​in the range (0, 1). k This signifies reducing the weight of σ(x) to shift the focus towards development;

[0058] The PI acquisition function is as follows:

[0059] PI estimates the function value using a posterior model, which is greater than the maximum value μ among the current actual observations. + The probability of finding the next sampling point is calculated using the following formula:

[0060]

[0061] Where Φ represents the normal cumulative distribution function, μ(x) and σ(x) represent the Gaussian mean and standard deviation of the posterior prediction, respectively, and ξ≥0 is an adjustable parameter, indicating that the objective is to find a value similar to μ. + The point with the highest probability of increasing ξ compared to the previous one;

[0062] The EI acquisition function is as follows:

[0063]

[0064]

[0065] The weights Φ(Z) are the normal cumulative distribution function, and ξ represents the relationship between μ and Z. + Compared to the degree of improvement, φ(Z) represents the normal probability distribution function. The goal of EI is to find the point with the maximum expected improvement, (μ(x)-μ + -ξ)Φ(Z) and σ(x)φ(Z) are used for development and exploration, respectively.

[0066] The beneficial effects of this invention are as follows:

[0067] 1) This invention adopts a closed-loop optimization experimental method, which uses the data from each optimization cycle to guide the next experiment, making full use of existing data. It does not require measuring all protocols in the entire charging protocol space, reducing the number of experimental tests and saving experimental time.

[0068] 2) This invention uses variance features that have a strong negative correlation with battery life to characterize lifespan. By using a Bayesian optimization model, the protocol with the smallest log-variance feature value is found to find the charging protocol with the longest lifespan. Only the data from the first n cycles of the battery need to be reduced, without having to cycle the battery to the end of its lifespan. This greatly reduces the number of experiments and saves optimization time. Lifespan is characterized by features without the need for actual lifespan measurement.

[0069] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0071] Figure 1 This is a schematic diagram illustrating the design process of the closed-loop optimization battery fast charging strategy based on feature-driven and machine learning as described in this invention.

[0072] Figure 2 The variance characteristics are based on the first 100 cycles, where (a) is the discharge capacity curve for the 100th and 10th cycles, (b) is the relationship curve between the difference in discharge capacity between the 100th and 10th cycles and voltage, (c) represents the correlation coefficient between the logarithmic variance of 45 batteries and cycle life, and (d) is the correlation coefficient between the logarithmic variance of 110 batteries and cycle life.

[0073] Figure 3 A schematic diagram of the Bayesian optimization process. Detailed Implementation

[0074] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0075] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0076] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0077] The space of fast charging protocols, which includes multiple constant current (CC) steps or constant voltage (CV) steps, is multidimensional. This requires measuring all protocols within the entire charging protocol space, and each experiment takes a lot of time. This invention uses a black-box optimized machine learning method to implement a closed-loop optimized experimental design (COED) method, which can use the data from each optimization loop to guide the next experiment, making full use of existing data and reducing the number of experimental tests.

[0078] The COED algorithm balances the conflict between exploration and exploitation. Exploration refers to exploring the untested parameter space with high uncertainty, while exploitation refers to testing promising regions based on the results of completed experiments. For example... Figure 1 As shown, the optimization system of the COED algorithm consists of three parts: a parallelizable battery experimental setup, a feature extractor, and a black-box optimization model. The parallelizable battery experimental setup cycles the battery for n cycles and acquires the cyclic data to be sent to the feature extractor. The smaller the number of cycles n, the less time the experiment takes. The value of n is related to the selected feature; in this embodiment, n = 100 cycles is selected.

[0079] Before optimization, m protocols are randomly selected from the entire protocol space. Based on the cyclic data of the selected protocols, the feature extractor passes features strongly related to battery life (e.g., features extracted from voltage, current, capacity, and temperature-time curves) to the black-box optimization model. This strong correlation can be linear, exponential, or logarithmic. After balancing exploration and development, the black-box optimization model recommends an appropriate charging protocol for each round. The recommended protocol is tested by the battery cycling component, and features are extracted from it and passed back to the black-box optimization model for iterative optimization. After a certain number of optimization cycles, the final recommended charging protocol is obtained.

[0080] The parameter space of the charging protocol used in this invention consists of 224 different six-step, ten-minute fast charging protocols. The charging protocol in this dataset aims to charge the battery's state of charge (SOC) from 0% to 80% in 10 minutes across j steps. j can be any integer value; in this embodiment, j = 4. Therefore, the battery's SOC is charged from 0% to 80% in four constant-current (CC) stages, totaling 10 minutes. The charging rates within the four CC stages are represented as CC1, CC2, CC3, and CC4, where 1C represents a full charge in one hour, and 2C represents a full charge in 0.5 hours. CC4 is determined by CC1, CC2, CC3, and the charging time (10 minutes) to reach 80% SOC. After the four constant-current (CC) stages, there is one more constant-current (CC) and constant-voltage (CV) charging stage, for a total of six stages ensuring the battery SOC reaches 100%. Since each step currently maintains 20% SOC, we can obtain:

[0081]

[0082] CC4 can be obtained under the constraint of charging time ranging from 0 to 80% in less than 10 minutes, as shown in the following formula:

[0083]

[0084] A total of 224 charging protocols are defined from CC1 to CC3. Each optimized cycle's recommended charging protocol is tested for 100 charge-discharge cycles using a battery testing setup. The cutoff current for the final constant-voltage charging during the charging process is C / 20. After a 5-second rest period at the end of charging, the discharging process begins, with all batteries discharging to 2V at a 4C discharge rate.

[0085] Features strongly correlated with battery life include voltage, current, capacity, and temperature-time curves. Finding a feature with a clear and strong correlation to cycle life can avoid building a life predictor and reduce overall optimization time. This invention seeks features from the battery's discharge capacity-voltage curve ΔQ(V) or voltage-discharge capacity curve ΔV(Q). Variance, mean square, kurtosis, skewness, maximum, and minimum value features can be obtained from both the ΔQ(V) and ΔV(Q) curves.

[0086] In this embodiment, the voltage versus discharge capacity curves ΔQ from the 100th and 10th cycles are used. 100-10 Explore features in (V), such as Figure 2 As shown in (a), after a certain number of cycles, the battery's discharge capacity at the same voltage will decrease, indicating that electrochemical information can be captured by the discharge voltage curve. Figure 2 (b) Indicates the ΔQ of different batteries 100-10 The (V) curve is related to the cycle life of a battery. Different curves show significant differences in the middle section, with the first and last parts almost overlapping or having slight variations. A longer battery life is indicated when the middle section is further to the right. This curve consists of a series of points with equal voltage intervals. The variance of the discharge capacity at these points can be used to characterize the degree of deviation in the middle section, and thus, the battery's lifespan.

[0087] Variance and battery life have a clear linear relationship in a logarithmic coordinate system, such as... Figure 2 As shown in (c) and (d). Figure 2 The data used in (c) corresponds to 45 batteries. Figure 2 The data used in (d) corresponds to 110 batteries.

[0088] from Figure 2 As can be seen in (c) and (d), there is a negative linear correlation between the logarithm of variance and the logarithm of lifetime and cycle lifetime, with correlation coefficients of -0.83 and -0.88, respectively. Figure 2 This indicates that using log(Var(Q)) 100-10 Using (V))) to characterize battery life is reasonable. Variance Var(Q) 100-10 Log(V)) of (V) 100-10 (V))) contains rich mechanistic information as the output and lifetime metric of Gaussian process regression (GPR), avoiding the need to obtain actual cycle lifetime through experiments that cost a lot of time and money, or to obtain predicted cycle lifetime using a trained lifetime predictor.

[0089] The black-box optimization in this invention can be a network search method, a random search method, Bayesian optimization, or other black-box optimization algorithms. In this embodiment, Bayesian optimization is used for prediction. Figure 3 As shown, Bayesian optimization mainly consists of Gaussian Process Regression (GPR) and a sampling function. This algorithm is based on GPR and uses sampling points to predict the probability distribution of function values ​​at each point. GPR obtains the output y and the standard deviation of all its inputs x through a probabilistic surrogate model and training data from experiments. The sampling function determines where to sample next based on the probabilistic surrogate model.

[0090] The Bayesian optimization process is as follows: First, randomly select several protocols (e.g., 48) from all 224 protocols and obtain their corresponding log-variance values. At this point, the training set for GPR is obtained. This training set takes the three-dimensional protocols (CC1-CC3) as input and the log-variance from the feature extractor as Log(Var(Q)). 100-10 (V))) is used as the output. A Gaussian model is trained using the training set, and then the log-variance (Log(Var(Q))) is predicted for all 224 protocols. 100-10 (V)). Using GPR, in addition to obtaining the mean, we can also obtain the standard deviation of each protocol, which is unavailable through other regression methods. Then, a sampling function is used to determine which protocols to test in the next round. The newly obtained protocols and their log-variance features obtained from the first 100 iterations are added to the original training set to obtain a new training set. This new training set is then used to retrain the Gaussian model and predict the log-variance Log(Var(Q)) again for all 224 protocols. 100-10 (V))) and then use the acquisition function to recommend a new protocol. Repeat the above process until the optimization objective is achieved, such as finding a protocol with a high cycle lifetime (corresponding to a low log-variance Log(Var(Q100-10(V)))).

[0091] The goal of Gaussian Process Regression (GPR) in this invention is to create a function mapping f: R3 → R using training data without needing to know the specific function expression. The input x of GPR is CC1-CC3, and the output y is the log-variance Log(Var(Q)). 100-10 (V))). Considering the noise at each point, the actual output y can be expressed as:

[0092] When applying GPR, the actual output y after considering noise is expressed as:

[0093] y i =f(x) i )+ε i

[0094] Where (x) i y iLet f represent the i-th data point, f(x) represent the feature map, and f(x) represent the i-th data point. i ) represents the output of the Gaussian model; ε i Indicates observation noise, and

[0095] The set of any finite inputs X = {x1, x2, ..., x...} n} d*n All of them follow a joint Gaussian probability distribution p(f(x1), f(x2), ..., f(x)). n The mean function and covariance function of f(x) are as follows:

[0096] m(x) = E[f(x)]

[0097] k(x,x′)=E[(f(x)-m(x))(f(x′)-m(x′))]

[0098] Due to the lack of prior information, m(x) is usually set to 0, and thus the distribution of y is written as:

[0099]

[0100] Where K ff Let be an N*N dimensional matrix, and [K ff ] i,j =k(x i ,x j );

[0101] For a new input Its joint distribution with the known observed target values ​​can be written as:

[0102]

[0103] in [K ff ] i,j =k(x i ,x j );

[0104] Conditional distribution p(f) * |X,y,X * In X * The formula is as follows:

[0105]

[0106] in, Let cov(f) represent the mean of the posterior distribution. * The variance of the posterior distribution is represented by the following formula:

[0107]

[0108]

[0109] The kernel functions that can be used in the GPR model include, but are not limited to, the following:

[0110] The first kernel function is the Gaussian kernel RBF, given by the following formula:

[0111]

[0112] Where |xx′| 2 Let σ represent the square of the magnitude of vector xx′, and let σ represent the standard deviation.

[0113] The second kernel function is the Matern kernel, given by the following formula:

[0114]

[0115] It has positive parameters v and l, and the possible values ​​of υ are 0.5, 1.5, and 2.5. Here, Γ(v) is a Γ function, Kν is a Bessel function of the second kind, and d(x,x′) represents the Euclidean distance;

[0116] The third kernel function is the white noise kernel function:

[0117]

[0118] Where, σ m is a constant representing the level of white noise.

[0119] 1) UCB (Upper Confidence Bound, UCB) data acquisition function:

[0120] UCB(x) = μ(x) + βσ(x)

[0121] Where μ(x) and σ(x) are the mean and standard deviation of the posterior distribution, respectively, and β is a hyperparameter to be optimized to balance exploration and development. Exploration refers to exploring untested parameter spaces with high uncertainty, while development refers to testing promising regions based on completed experimental results. As the number of steps increases, the focus shifts from exploration to development, and β is gradually reduced, resulting in:

[0122] UCB(x)=μ(x)+β0ε k σ(x)

[0123] Where β0,ε k All are hyperparameters, k represents the number of optimizations, and ε takes values ​​in the range (0,1). k This signifies reducing the weight of σ(x) to shift the focus towards development;

[0124] 2) PI (Probability of Improvement) Acquisition Function:

[0125] PI estimates the function value using a posterior model, which is greater than the maximum value μ among the current actual observations. + The probability of finding the next sampling point is calculated using the following formula:

[0126]

[0127] Where Φ represents the normal cumulative distribution function, μ(x) and σ(x) represent the Gaussian mean and standard deviation of the posterior prediction, respectively, and ξ≥0 is an adjustable parameter, indicating that the objective is to find a value similar to μ. + The point with the highest probability of increasing ξ compared to the previous one;

[0128] 3) Expected Improvement (EI):

[0129]

[0130]

[0131] The weights Φ(Z) are the normal cumulative distribution function, and ξ represents the relationship between μ and Z. + Compared to the degree of improvement, φ(Z) represents the normal probability distribution function. The goal of EI is to find the point with the maximum expected improvement, (μ(x)-μ + -ξ)Φ(Z) and σ(x)φ(Z) are used for development and exploration, respectively.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A closed-loop optimization method for designing fast-charging strategies for batteries based on feature-driven and machine learning, characterized in that: Includes the following steps: S1: Based on the closed-loop optimization experimental design method COED, randomly select from all charging protocols. m A charging protocol, using a battery testing device to test the battery. n One charge-discharge cycle; S2: During the loop, use the feature extractor from the previous... n Extracted from the next charge-discharge cycle m Features related to battery life corresponding to each charging protocol are used as the output of the training set of the black-box optimization model, and together with the charging protocols used as input, they form the training set to train the black-box optimization model. S3: Utilize the trained black-box optimization model to predict the features of all charging protocols and obtain the protocols to be tested in the next round. S4: Add the protocol to be tested in the next round and its battery life-related features to the original training set to obtain a new training set to retrain the black-box optimization model. Repeat steps S3-S4 until the conditions are met or the optimization goal is achieved. The feature extractor extracts m The feature corresponding to the charging protocol is the variance feature, and the first charging protocol is the first charging protocol. n Second and third p The logarithm of the variance of the difference between the discharge capacity and voltage curves over several charge cycles is used as a characteristic of battery life and is expressed as the log-variance feature. Log ( Var ( Q n-p ( V ))),in p Less than n Positive integers; The black-box optimization model is a Bayesian optimization model, which consists of a Gaussian process regression model (GPR) and a data acquisition function. In step S2, m A three-dimensional charging protocol and its corresponding log-variance characteristics Log ( Var ( Q n-p ( V The GPR model is trained using the training set to output the log-variance features of all three-dimensional charging protocols. The acquisition function is used to determine which protocols to test in the next round; The GPR is used to determine the feature values ​​corresponding to each protocol. The input is... The three-dimensional protocol outputs log-variance features. Log ( Var ( Q n-p ( V ))); When applying GPR, the actual output after considering noise. y Represented as: in( ) indicates the first i Data points, Represents feature mapping, This represents the output of the Gaussian model; Indicates observation noise, and ; A set of any finite input They all follow a joint Gaussian probability distribution. , The mean function and covariance function are as follows: Because there is no Prior information about the distribution, Set to 0, and then y The distribution is written as: in for N N A matrix of dimension, and ; For a new input The joint distribution of this distribution and the known observed target values ​​can be written as: in , ; Conditional distribution Follows a normal distribution: in, Let represent the mean of the posterior distribution. The variance of the posterior distribution is expressed as follows: 。 2. The closed-loop optimization battery fast charging strategy design method based on feature-driven and machine learning according to claim 1, characterized in that: All the charging protocols constitute a charging protocol space. The charging protocols within this space are designed to charge the battery's State of Charge (SOC) from 0% to 80% within 10 minutes using multiple stages of constant current charging. During this process, the following methods are used: j Different currents of varying magnitudes are used to charge the battery in stages.

3. The closed-loop optimization battery fast charging strategy design method based on feature-driven and machine learning according to claim 2, characterized in that: The charging protocol is Multi-Stage Constant Current Constant Voltage (MSCC), which consists of multiple constant current charging stages followed by a constant voltage charging stage. when j When the value is 3, the charging protocol charges the battery from 0% to 80% SOC in 3 steps within 10 minutes. The parameters of the charging protocol are: , This is also the two-dimensional input of the black-box optimization model; specifically: the charging rate in the 0-26.67% SOC range. The charging rate in the 26.67%-53.34% SOC range is The charging rate in the 53.34%-80% SOC range is ,in Size by , A decision is expressed as: when j When the value is 4, the charging protocol charges the battery from 0% to 80% SOC in 4 steps within 10 minutes. The parameters of the charging protocol are: This is also the three-dimensional input of the black-box optimization model; specifically: the charging rate in the 0-20% SOC range. The charging rate in the 21%-40% SOC range is The charging rate in the 41%-60% SOC range is The charging rate in the 61%-80% SOC range is ,in Size by The decision is made using the following formula: when j When the battery is at =5, the charging protocol charges the battery from 0% to 80% SOC in 5 steps within 10 minutes. The parameters of the charging protocol are: , , , This is also the four-dimensional input of the black-box optimization model; specifically: the charging rate in the 0-16% SOC range. The charging rate in the 16%-32% SOC range is The charging rate in the 32%-48% SOC range is The charging rate in the 48%-64% SOC range is The charging rate in the 64%-80% SOC range is ,in Size by , , , The decision is made using the following formula: 。 4. The closed-loop optimization battery fast charging strategy design method based on feature-driven and machine learning according to claim 1, characterized in that: The kernel functions used in the GPR model include the following: The first kernel function is the Gaussian kernel RBF, given by the following formula: in Representing vectors The square of the modulus, Indicates standard deviation; The second kernel function is The core is given by the following formula: It has a positive parameter and ,in It is function, It is a Bessel function of the second kind. Indicates Euclidean distance; The third kernel function is the white noise kernel function: in, is a constant representing the level of white noise.

5. The closed-loop optimization battery fast charging strategy design method based on feature-driven and machine learning according to claim 1, characterized in that: The acquisition function is the confidence interval upper bound function UCB, the improved probability function PI, or the expected increment function EI, where: The UCB acquisition function is as follows: in and These are the Gaussian mean and standard deviation of the posterior predictions, respectively. This is a hyperparameter that needs optimization to balance exploration and development. Exploration involves exploring an untested parameter space with high uncertainty, while development involves testing promising regions based on completed experimental results. As the number of steps increases, the focus shifts from exploration to development, gradually reducing... ,get: in All of these are hyperparameters. Indicates the number of optimizations. The value range is (0,1). Then the characterization decreases The weighting of this will shift the focus to development; The PI acquisition function is as follows: The PI estimates the function value using a posterior model, which is greater than the maximum value among the current actual observations. The probability is used to find the next sampling point. The probability is expressed by the following formula: in Φ Represents the normal cumulative distribution function. It is an adjustable parameter, indicating that the goal is to find the match between... improvement The point with the highest probability; The EI acquisition function is as follows: Weight It is the normal cumulative distribution function. Let represent the normal probability distribution function. The goal of EI is to find the point with the maximum expected improvement. and They are used for development and exploration, respectively.

Images