Real-time estimation method of lithium-ion battery capacity based on canonical correlation analysis

Abstract

The present invention discloses a real-time estimation method for the power of a lithium-ion battery based on canonical correlation analysis, which belongs to the technical field of electric vehicle battery management and relates to a method for estimating the power of a lithium-ion battery. The method uses battery current, voltage, and temperature data collected under different battery operating conditions to calculate the true power value as output label data; a real-time power estimation model is constructed by regressing the power through canonical correlation analysis and training the model parameters to obtain a trained prediction model, in which the model input is battery characteristic data and the output is the power value to be estimated; based on the real-time estimation of the battery status, the power estimation value is updated under a given noise situation through particle filtering to obtain the optimal power estimation.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric vehicle battery management and relates to a method for estimating the power of a lithium-ion battery. Specifically, it relates to a method for estimating the power of an electric vehicle lithium-ion battery based on offline training of canonical correlation analysis and online error update of particle filtering. The method can improve the accuracy of lithium-ion battery power estimation, reduce computing consumption, and expand the scope of application. Background Art

[0002] The battery management system (BMS) is a crucial component of electric vehicles, encompassing functions such as state of charge (SOC) estimation, state of health (SOH) estimation, and pack balancing. SOC estimation is a crucial and fundamental function. It calculates the remaining battery charge percentage in real time by monitoring data such as current, voltage, and temperature. This information is then communicated to the user by the BMS, impacting the electric vehicle's current remaining charge and remaining range. Therefore, designing SOC estimation methods for BMSs and optimizing the calculations to improve accuracy is crucial.

[0003] Existing state-of-charge estimation methods for battery system design are primarily based on an equivalent circuit model of the battery, using Kalman filtering to achieve real-time estimation. Existing battery charge estimation models are traditionally empirical models that require engineers to have extensive engineering experience. Furthermore, Kalman filtering assumes that the noise is Gaussian, whereas in reality, the noise may be non-Gaussian. In recent years, with the advancement of computer technology and the continuous improvement of scientific computing capabilities, data-driven methods have been widely used. These methods can avoid the complexity of models and parameter uncertainty, requiring only training data through methods such as neural networks to obtain charge predictions. However, not only do existing model methods have limitations, but data-driven methods also suffer from problems such as overly complex neural networks. The number of network layers and parameters are difficult to determine, making it impossible to obtain a unified, concise algorithm model, and the estimation accuracy is poor. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a battery power estimation method based on offline training of canonical correlation analysis and online update error of particle filtering, which realizes power estimation based on offline data training and online update error, can improve the accuracy of lithium battery power estimation, reduce computational consumption, and has universal applicability.

[0005] The present invention first collects the current, voltage, and temperature data of the battery under different operating conditions in the laboratory, and performs offline data training based on canonical correlation analysis to obtain the regression relationship between the current, voltage, and temperature characteristic data and the remaining power. Then, based on the regression parameters obtained from the training, the power estimation value under the new data to be tested is predicted. Finally, this estimation value is combined with the Coulomb integral formula model to perform filter update estimation under the particle filter framework to reduce the estimation error. The innovation of the present invention lies in proposing a regression model based on canonical correlation analysis, and combining it with particle filtering to obtain online power estimation updates, so that the data can obtain the optimal power estimation result under the condition of uncertain noise. The existing power estimation method obtains the power estimation result based on the equivalent circuit model and its estimated parameters and Kalman filtering; while the present invention, based on considering the correlation between the current, voltage, and temperature data and the power to be estimated, proposes an easy-to-implement offline data regression training method, obtains a preliminary power estimation based on the canonical correlation analysis regression model, and considers that the actual noise may be non-Gaussian noise, and obtains the power filter update estimate based on the particle filter that can realize any noise condition. In practical applications, the present invention does not have high computational performance for the battery management system, because the data training proposed by the present invention is an offline process that can be performed in the laboratory. It only requires exporting the training parameters to the battery management system and combining them with the real-time current state to predict the optimal filter estimate.

[0006] The battery power estimation method based on canonical correlation analysis and particle filtering proposed in the present invention can simplify calculations and improve the accuracy of power estimation. The principle is: because the parameters to be identified in the traditional model cannot be accurately obtained and change dynamically, on the other hand, the methods based on neural networks are usually complex in structure and have many parameters, and the present invention uses canonical correlation analysis for data training, without the need to establish a mechanism model. This first solves the problem of difficulty in parameter identification of the traditional model, and secondly avoids the problem of complex structure of deep neural networks. This process is the first approach of the present invention. The present invention considers canonical correlation analysis regression prediction of power estimation, and further particle filtering is performed based on the regression estimate to obtain the optimal power estimation under noise conditions. The power regression estimation method proposed in the present invention is essentially different from the existing methods. The existing methods mainly focus on the structural adjustment and hyperparameter setting of the neural network model, but the only deep network cannot obtain online error updates. The method of the present invention includes prediction and update processes, which can achieve real-time estimation of battery power.

[0007] The technical solution provided by the present invention is:

[0008] A real-time estimation method for the lithium battery capacity of electric vehicles based on canonical correlation analysis and particle filtering is a data-driven capacity regression estimation prediction model method. It uses the battery current, voltage, and temperature collected under different battery operating conditions as input data, obtains the true capacity value through Coulomb calculation, and uses it as output label data. By constructing a real-time capacity estimation model with canonical correlation analysis regression and training the model parameters, a trained prediction model is obtained (the model input is the battery characteristic data and the output is the value to be estimated). In actual use, the method estimates the battery status in real time, and combines particle filtering to update the capacity estimation value under given noise conditions to obtain the optimal capacity estimation. It includes offline and online processes. The offline process includes data-driven canonical correlation analysis regression estimation, and the online process includes particle filtering under noise conditions. It mainly includes the following steps:

[0009] 1) Perform canonical correlation modeling and establish a canonical correlation coefficient, which reflects the correlation between battery input characteristics and output power. Input data includes battery current, voltage, and temperature data, and the output is label data of the power state.

[0010] 2) Collect battery charge and discharge data under dynamic conditions and perform data preprocessing;

[0011] To adapt to the actual working state of the battery, the present invention collects battery charge and discharge data under dynamic working conditions for offline training. The current, voltage, and temperature data are used as input, and the coulomb calculation formula is used to obtain the state of charge as the label output. The training process using raw data requires data preprocessing, including the following steps:

[0012] 21) Select current, voltage, and temperature as input feature data to construct an input matrix; select the power state as output to construct an output data matrix;

[0013] 22) Mapping the input feature data to a feature vector through a polynomial to obtain feature dimension expansion;

[0014] Collect all the time data and stack them to get the input matrix Contains N rows, representing t1 to t N moment; for the input matrix Each row of data is polynomially mapped to obtain the expanded input matrix, which is expressed as:

[0015]

[0016] in, is the expanded input matrix; the battery characteristic input variable a∈R at each moment p , p represents a p-dimensional vector, and each row represents the input feature data of a moment of dimension expansion, such as [φ1(a(t1)) φ2(a(t1)) …φ n(a(t1))] represents the expanded input feature data at time t1.

[0017] The power state output variable b∈R q ; Obtain the true state of charge label b = [SOC] at each moment through the Coulomb calculation formula; Construct the output data matrix:

[0018]

[0019] Each row represents the actual power status label data at each moment, containing N rows, representing t1 to t N time.

[0020] 23) Perform data normalization;

[0021] Expand the matrix of battery input feature data Each row is normalized and preprocessed, expressed as:

[0022]

[0023] The operation is calculated element by element on the matrix, μ1 represents the battery feature input data expansion matrix The mean vector of the column-by-column elements, σ1 represents the input data expansion matrix The column-wise standard deviation vector of .

[0024] Also output data matrix for power label Each row is normalized and expressed as:

[0025]

[0026] The operation is calculated element by element on the matrix, μ2 represents the power label output data matrix The mean vector of the elements of each column, σ2 represents the output matrix The column-wise standard deviation vector of .

[0027] 24) Divide the data into a training set and a cross-validation set;

[0028] 3) Establish a real-time estimation model of power consumption based on correlation analysis and regression;

[0029] The objective function of the correlation analysis regression real-time power estimation model is expressed as:

[0030]

[0031] Where A and B are the battery input data and battery output data preprocessed in step 2); β is the typical correlation parameter to be optimized for the input matrix A, and θ is the typical correlation parameter to be optimized for the output matrix B; β∈R p,θ∈R q ; λ1 is the regularization penalty hyperparameter for the regression parameter β term, and λ2 is the regularization penalty hyperparameter for the regression parameter θ term, which is given in advance and used to optimize the established correlation analysis regression model; ‖‖ 2 represents the square of the l2 norm;

[0032] 4) Solving the model to obtain parameters of a real-time estimation model for canonical correlation analysis and regression; estimating the real-time remaining battery capacity based on the parameters of the solved canonical correlation analysis and regression model;

[0033] The real-time power estimation model solution includes the following steps:

[0034] 41) Adjust and determine the values of hyperparameters, including λ1, λ2 and the degree of polynomial mapping;

[0035] 42) Under given hyperparameter values, optimize the objective function of canonical correlation analysis (Equation (11)). Use the gradient descent method with the help of the fmincon function in the MATLAB toolkit to solve the constrained minimization objective function. Obtain the prediction parameters of the canonical correlation analysis model (canonical correlation analysis regression parameters β, θ);

[0036] 43) For the newly collected battery input feature data a = [I, V, T], the battery power state is calculated by regression, which is expressed as:

[0037]

[0038] in, is the battery charge status obtained by canonical correlation analysis regression.

[0039] 5) Combined with the particle filter online update error, the optimal state of charge estimate is obtained;

[0040] Through steps 1) to 5), a battery state of charge estimation method based on canonical correlation analysis offline training and particle filtering online update optimal estimation is realized.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention provides a method for estimating the battery state of charge based on data-driven canonical correlation analysis regression. The battery state of charge is predicted by solving parameters through a canonical correlation analysis regression offline training model, and is combined with a particle filter to form a canonical correlation analysis offline data training and particle filter online update state estimation. The characteristic data collected by the battery under complex dynamic working conditions is used as input, and the state of charge label value is used as output. According to the idea of maximizing the correlation between the two groups of data through canonical correlation analysis, a regression estimation model is derived, which reduces the computational complexity. Combined with the Coulomb calculation formula, the optimal state estimation update of arbitrary noise is realized in the particle filter, thereby improving the accuracy of the lithium battery state of charge estimation and expanding the applicability of the battery state of charge estimation method. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a general flow chart of a battery state estimation method based on canonical correlation analysis and particle filtering proposed by the present invention.

[0044] Figure 2 This is a flowchart of the data selection preprocessing provided by the present invention.

[0045] Figure 3 This is a schematic diagram of the canonical correlation analysis regression training prediction proposed in the present invention.

[0046] Figure 4 This is a flowchart of the filtering update and optimal state estimation process provided by the present invention.

[0047] Figure 5 This is a result diagram of the battery power estimation example provided by the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below by way of examples in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0049] The present invention provides a real-time estimation method for the electric vehicle lithium battery (lithium-ion battery) power based on canonical correlation analysis and particle filtering. The method is a data-driven power regression estimation prediction model method. The method uses battery current, voltage, and temperature collected under different battery operating conditions as input data, obtains the true power value through Coulomb calculation, and uses it as output label data. The model parameters are trained through canonical correlation analysis regression to obtain a prediction model of battery characteristic input data and output estimated values. In actual use, the method estimates the battery state in real time and combines particle filtering with a given noise situation to update the power estimate value to obtain the optimal power estimate. The method includes an offline process and an online process. The offline process includes data-driven canonical correlation analysis regression estimation, and the online process includes particle filtering under noise conditions. Figure 1 The overall process of the method of the present invention is shown; it includes the following steps:

[0050] 1) Perform canonical correlation modeling and establish a canonical correlation coefficient, which reflects the correlation between battery input characteristics and output. Input data includes battery current, voltage, and temperature data, and the output is the label data of the battery state.

[0051] In typical correlation modeling, it is assumed that the battery characteristic input variable a∈R at each moment p , power state output variable b∈R q , canonical correlation analysis is done by finding the canonical correlation parameter β∈R p ,θ∈R q , p and q are vector dimensions; further, two sets of linear combinations are obtained from the parameters and input and output data respectively:

[0052] U=β T a=β1a1+β2a2+…+β p a p ,

[0053] V=θ T b=θ1b1+θ2b2+…+θ q b q Formula (1)

[0054] Canonical correlation analysis seeks the solution of parameters β and θ so that the correlation coefficient of the above linear combination is maximized. The optimization form is:

[0055]

[0056] The present invention establishes the correlation between the battery feature input and the state of charge output by taking the input current, voltage, and temperature as input a and the power label data as output b, and maximizing the typical correlation coefficient. In Formula 2, ρ is the typical correlation coefficient, β is the typical correlation parameter about input a, θ is the typical correlation parameter about output b, and ∑ aa is the covariance matrix of feature input a, ∑ bb is the covariance matrix of the power state output b, ∑ ab is the covariance matrix between the feature input a and the power output b.

[0057] 2) Collect battery charge and discharge data under dynamic conditions and perform data preprocessing;

[0058] In order to adapt to the actual working state of the battery, the present invention collects the battery charge and discharge data under dynamic working conditions for offline training. The current, voltage and temperature data are used as input, and the coulomb calculation formula is used to obtain the state of charge as the label output. The training process using the original data requires data preprocessing. Figure 2 The method flow of data selection preprocessing is shown, which includes the following steps:

[0059] 21) Input and output data selection;

[0060] The battery charge and discharge data under dynamic working conditions are sampled at each moment t i The current I(t i ), voltage V(t i ), temperature T(t i ) as input a, that is, a(t i )=[I(t i ),V(t i ),T(t i )], the data sampling frequency is 1Hz, and the input matrix is constructed:

[0061]

[0062] Each row of Formula 3 represents the input feature data at a moment, such as a(t1) represents the input feature data at time t1; all the time data are collected and stacked to obtain the input matrix Contains N rows, representing t1 to t N Moment. The Coulomb calculation formula is used to obtain the value of t at each moment. i The real power state label b, that is, b(t i )=[SOC(t i )], construct the output data matrix:

[0063]

[0064] Each row of Equation 4 represents the actual state of charge tag data at the corresponding time of Equation 3, and also contains N rows, representing the time from t1 to t N time.

[0065] 22) Mapping the input feature data to a feature vector through a polynomial to obtain feature dimension expansion;

[0066] In order to better reflect the characteristics of the input data, the present invention uses polynomial mapping to map the input feature data to a vector in a higher dimensional space, namely The polynomial mapping φ maps the original vector to a higher-order polynomial of all elements to obtain more elements. n represents the number of all elements after mapping, which depends on the number of times the highest-order term of the mapping is. For example, a polynomial is mapped to the highest quadratic term, then The number of elements n after mapping is 9.

[0067] Perform polynomial mapping on each row of the input matrix of Formula 3, and the expanded data is:

[0068]

[0069] 23) Data normalization;

[0070] In order to make different elements of the data in a similar range to prevent overfitting or underfitting in the training process, the battery input feature data is expanded Each row is normalized and preprocessed, which can be expressed as:

[0071]

[0072] The operation is calculated element by element on the matrix, μ1 represents the battery feature input data expansion matrix The mean vector of the column-by-column elements, σ1 represents the input data matrix The column-wise standard deviation vector of .

[0073] Similarly, for the power label data matrix Each row of is normalized and can be expressed as:

[0074]

[0075] The operation is calculated element by element on the matrix, μ2 represents the mean vector of the column-by-column elements of the power label output data matrix B, and σ2 represents the output matrix The column-wise standard deviation vector of .

[0076] 24) The data is divided into a training set and a cross-validation set;

[0077] After all the data are obtained, all the data are randomly divided into 80% for training data set and 20% for cross-validation set. The purpose of dividing the data into training set and validation set is to train the data through the training set, and use the results for cross-validation set prediction to prevent overfitting of the training data, and thus determine the choice of hyperparameters.

[0078] 3) Establishing a canonical correlation analysis regression model for real-time power estimation, and solving the model to obtain the parameters of the canonical correlation analysis regression model;

[0079] In step 2), the data has been normalized, and the covariance matrix of the battery feature input and the power state output can be written as The covariance matrix of the battery feature input can be written as The covariance matrix of the state of charge output can be written as In order to obtain the regression form of canonical correlation analysis, the canonical correlation analysis optimization formula (2) is written as the optimization objective function in the following form:

[0080]

[0081] Among them, β is the typical correlation parameter about the input to be optimized, and θ is the typical correlation parameter about the output to be solved;

[0082] Adding a negative sign to the objective function of the above formula can transform the maximization problem into a minimization problem. At the same time, the denominator can be constrained to a unit value, so the above formula can be written as:

[0083]

[0084] The above formula can be equivalently written as follows:

[0085]

[0086] In order to avoid overfitting in the training process, the present invention adopts the general l2 norm to impose penalty constraints on the parameters. The optimization form after adding the norm penalty constraint is:

[0087]

[0088] The above formula is the objective function of the canonical correlation analysis regression model proposed in the present invention, where A and B are the input and output data preprocessed in step 2), λ1 is the regularization penalty term hyperparameter for the regression parameter β term, and λ2 is the regularization penalty term hyperparameter for the regression parameter θ term, which are given in advance and used to optimize the above model.

[0089] 4) Regress the parameters of the real-time power estimation model based on the obtained canonical correlation analysis to estimate the real-time remaining battery power;

[0090] Figure 3 This is the canonical correlation analysis regression training and prediction process proposed in the present invention. The canonical correlation analysis regression optimization objective (Equation (11)) used in the present invention includes a hyperparameter penalty term, and the traditional singular value decomposition method cannot directly obtain an analytical solution. To solve the optimization problem, the present invention uses the fmincon function of the MATLAB toolkit to solve the problem of minimizing the objective function under constraints.

[0091] Before training (solving) formula (11), it is necessary to give the values of the hyperparameters. In order to determine the final values of the hyperparameters, the present invention conducts multiple trainings, each time adjusting the values of different hyperparameters, including λ1, λ2 and the number of polynomial mappings. Under each given hyperparameter value, the objective function of the canonical correlation analysis (formula (11)) is optimized to obtain the prediction parameters of the canonical correlation analysis model (canonical correlation analysis regression parameters β, θ), and the final values of the hyperparameters are determined based on the prediction accuracy of the optimized training set and the cross-validation set under multiple different hyperparameter values.

[0092] In the specific implementation of the present invention, in addition to using the typical correlation coefficient to reflect the correlation of input and output data, a prediction evaluation index is also established to evaluate the prediction accuracy of the power state prediction value;

[0093] In order to quantitatively evaluate the quality of the prediction method, two evaluation indicators, root mean square error RMSE and mean absolute error MAE, are established:

[0094]

[0095]

[0096] Among them, SOC i is the actual state of charge value of the i-th data point, is its estimated value, and N is the sample size.

[0097] After determining the choice of hyperparameters, the objective function is optimized to solve the canonical correlation analysis regression parameters β and θ (canonical correlation analysis model prediction parameters). For the newly collected battery input feature data a = [I, V, T], the battery charge state can be regressed to obtain:

[0098]

[0099] in, It is the power state prediction value obtained by canonical correlation analysis regression.

[0100] 5) Combined with the particle filter online update error, the optimal state of charge estimate is obtained;

[0101] The present invention predicts the battery state of charge by solving parameters of a canonical correlation analysis regression offline training model, and combines it with particle filtering for online updating to obtain the optimal state of charge estimate under noise conditions. Figure 4 The figure is a flow chart of particle filtering. Particle filtering is a form of filtering used when both the state equation and the measurement equation are noisy. Unlike Kalman filtering, which is only applicable to linear systems and Gaussian noise, particle filtering can be applied to nonlinear systems and arbitrary noise. In the problem of battery state estimation, the noise of the dynamic equations and measurement equations that need to be considered is not necessarily Gaussian noise, so the use of particle filtering can be applied to this situation. The basic idea of particle filtering is to use a series of particles to represent the posterior estimate of the state. In the prediction of battery state of charge, consider the following state space description form:

[0102]

[0103] The first line is the discrete coulomb calculation form, which represents the actual state of charge SOC at time k. k and k-1 moment SOC k-1 The dynamic relationship between k-1Indicates the current at time k-1, which is positive during charging and negative during discharging. Δt is the sampling time interval of 1s. C represents the total capacity of the battery. r is the input noise, which can be any form of noise, not limited to Gaussian noise. k It represents the preliminary estimated value of the battery state of charge at the kth moment obtained by using canonical correlation analysis regression in the process 1)-4). Estimated and true SOC k There is an estimation error between them, which is expressed as noise v.

[0104] The battery dynamic system equation (13) can be written as

[0105] x k =g k (x k-1 ,r k-1 )

[0106] y k =h k (x k ,v k ) Formula (14)

[0107] In this step x k Expressed as the system state, that is, the real state of charge SOC in formula (13) k ,y k is the system output, that is, CCA in formula (13) k , g k is the dynamic equation expressed by formula (13), h k is the output equation expressed by equation (13). To solve a state f(x k ) is estimated a posteriori using importance sampling The posterior estimate expectation of the state is The weight of the i-th particle It can be written iteratively as:

[0108]

[0109] The process of particle filtering is:

[0110] First, initialize the initial states of M particles at the initial moment The particles can be set around an initial estimate of the battery charge based on known information.

[0111] Then the sampling hypothesis distribution Set to the distribution corresponding to the dynamic equation Then the weight in formula (15) can be written as:

[0112]

[0113] To avoid the phenomenon of particle degradation where after several iterations only a few particles have large weights while the weights of most other particles are close to 0, a particle resampling process is used:

[0114]

[0115] After resampling, all particles have the same weight The posterior estimate of the battery state of charge at time k is:

[0116]

[0117] By repeating the above process, input data can be continuously measured during the battery operation process, and a preliminary state of charge estimate can be obtained through canonical correlation analysis regression, and then the optimal state of charge estimate under noise can be obtained through particle filtering.

[0118] Through steps 1) to 5), a battery state of charge estimation method based on canonical correlation analysis offline training and particle filtering online update optimal estimation is realized.

[0119] Algorithm 1: Offline training of canonical correlation analysis to solve the canonical correlation analysis regression parameters, including:

[0120] 1. Collect historical data and construct the battery feature input data matrix A and the power label output matrix B;

[0121] 2. Perform polynomial mapping and normalization on the data, and divide it into training set and cross-validation set;

[0122] 3. Optimize the objective function of the canonical correlation analysis regression model and solve the canonical correlation analysis regression parameters β and θ:

[0123]

[0124] 4. Predict the state of charge under the new feature data input a = [I, V, T]:

[0125]

[0126] Algorithm 2 combines the Coulomb calculation formula with particle filtering to obtain an updated state of charge estimate, including:

[0127] 1. Using the canonical correlation analysis regression results as the output equation and the Coulomb calculation as the dynamic equation, the dynamic system form of the battery under noise data is constructed:

[0128]

[0129] 2. Initialize particles;

[0130] 3. Sample the particles, calculate the weights, resample, and calculate the optimal battery state of charge estimate.

[0131] Figure 5 The battery capacity estimation example results provided by the present invention are as follows: Figure 5 As shown, a certain brand of 18650 model ternary lithium battery was used in the experiment, and a section of battery data from 80% to 0% battery state under dynamic working conditions was selected. The solid line in the figure represents the actual state of charge label value, and the dotted line is the result of the prediction and estimation of the present invention. The calculation process is divided into two parts. First, according to Algorithm 1, the canonical correlation analysis regression model of the battery under a large amount of data is trained offline to obtain the canonical correlation analysis regression parameters, and the state of charge estimation result under this section of new data is calculated. Secondly, according to Algorithm 2, the canonical correlation analysis regression model is combined with Coulomb calculation to obtain the dynamic system equation of the battery under noise conditions. The error is updated using particle filtering to obtain an updated optimal state of charge estimation result.

[0132] It should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. However, those skilled in the art will appreciate that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A real-time lithium-ion battery capacity estimation method based on canonical correlation analysis uses battery current, voltage, and temperature data collected under different battery operating conditions to calculate the actual capacity value as output label data; A real-time power estimation model based on canonical correlation analysis and regression is constructed and the model parameters are trained to obtain a trained prediction model. The model input is battery characteristic data and the output is the power value to be estimated. Based on the real-time battery status estimation, the power estimation value is updated under given noise conditions through particle filtering to obtain the optimal power estimation. The model includes the following steps: 1) Collect battery current, voltage, and temperature data under different battery operating conditions; perform canonical correlation modeling and establish a canonical correlation coefficient to reflect the correlation between battery input characteristics and output power; the input includes battery current, voltage, and temperature data, and the output is label data of the power state; The canonical correlation modeling is performed to establish the correlation between the battery feature input and the state of charge output by maximizing the canonical correlation coefficient, including: Assume that the battery characteristic input variable a∈R at each moment p , power state output variable b∈R q , battery characteristics include current, voltage, and temperature; the state of charge is the charge label data; through canonical correlation analysis, find the typical correlation parameter β∈R to be solved p ,θ∈R q , p and q are vector dimensions; further, two sets of linear combinations are obtained from the parameters and input and output data, respectively, expressed as: U=β T a=β1a1+β2a2+…+β p a p , V=θ T b=θ1b1+θ2b2+…+θ q b q formula(1) Canonical correlation analysis seeks the solution of parameters β and θ so that the correlation coefficient of the above linear combination is maximized. The optimization form is expressed as: Where ρ is the typical correlation coefficient, β is the typical correlation parameter about the characteristic input variable a, θ is the typical correlation parameter about the power state output b, ∑ aa is the covariance matrix of feature input a, ∑ bb is the covariance matrix of the power state output b, ∑ ab is the covariance matrix between feature input a and power output b; 2) Perform data preprocessing, construct the battery input data matrix and the battery output data matrix, and divide the data into a training set and a cross-validation set; 3) Establish a real-time estimation model of power consumption based on correlation analysis and regression; The objective function of the correlation analysis regression real-time power estimation model is expressed as: Where A and B are the battery input data matrix and battery output data matrix preprocessed in step 2); β is the typical correlation parameter to be optimized for the input matrix A, and θ is the typical correlation parameter to be optimized for the output matrix B; β∈R p ,θ∈R q ; λ1 is the regularization penalty hyperparameter for the regression parameter β term, and λ2 is the regularization penalty hyperparameter for the regression parameter θ term, which are used to optimize the established correlation analysis regression model; ‖‖ 2 represents the square of the l2 norm; 4) Perform correlation analysis and regression to solve the real-time battery power estimation model and obtain model parameters; and estimate the real-time remaining battery power based on the solved model parameters; The real-time power estimation model solution includes the following steps: 41) Adjust and determine the values of hyperparameters, including λ1, λ2 and the degree of polynomial mapping; 42) Under given hyperparameter values, optimize the objective function of canonical correlation analysis and use the gradient descent method with the help of the fmincon function of the MATLAB toolkit to solve the constrained minimization objective function; obtain the parameters β and θ; The newly collected battery input feature data a = [I, V, T] is used as the data to be predicted. The battery power state is obtained through regression, which is expressed as: in, is the battery state of charge obtained by canonical correlation analysis regression; μ1 represents the mean vector of the elements of the battery feature input data expansion matrix column by column, σ1 represents the standard deviation vector of the elements of the input data expansion matrix column by column; φ represents the polynomial mapping; 5) Update the error online through particle filtering to obtain the optimal state of charge estimate; Through the above steps, real-time estimation of lithium-ion battery capacity based on canonical correlation analysis is achieved.

2. The method for real-time estimation of lithium-ion battery capacity based on canonical correlation analysis according to claim 1, wherein: Step 2) Data preprocessing includes the following steps: 21) Select current, voltage, and temperature as input feature data to construct an input matrix; use the state of charge as output to construct an output data matrix; 22) Mapping the input feature data to a feature vector through a polynomial to obtain feature dimension expansion; Collect all the time data and stack them to get the input matrix Contains N rows, representing t1 to t N At this moment, perform polynomial mapping on each row of the input matrix to obtain the expanded input matrix, which is expressed as: in, is the expanded input matrix; the battery characteristic input variable a∈R at each moment p , p represents a p-dimensional vector, and each row represents the input feature data of the expanded dimension at a moment; [φ1(a(t1)) φ2(a(t1)) … φ n (a(t1))] represents the expanded input feature data at time t1; The power state output variable b∈R q ; Obtain the true state of charge label b = [SOC] at each moment through the Coulomb calculation formula; Construct the output data matrix: Each row represents the actual power status label data at each moment, containing N rows, representing t1 to t N time; 23) Perform data normalization; Expand the matrix of battery input feature data Each row is normalized and preprocessed, expressed as: The operation is calculated element by element on the matrix, μ1 represents the battery feature input data expansion matrix The mean vector of the column-by-column elements, σ1 represents the input data expansion matrix The column-wise standard deviation vector of ; Also output data matrix for power label Each row is normalized and expressed as: The operation is calculated element by element on the matrix, μ2 represents the power label output data matrix The mean vector of the elements of each column, σ2 represents the output matrix The column-wise standard deviation vector of ; 24) Divide the data into a training set and a cross-validation set.

3. The method for real-time estimation of lithium-ion battery capacity based on canonical correlation analysis according to claim 1, wherein: The prediction evaluation indicators for evaluating the prediction accuracy of the power state prediction value are established, including the root mean square error RMSE and the mean absolute error MAE, which are expressed as: Among them, SOC i is the actual state of charge value of the i-th data point, is its estimated value, and N is the sample size.

Images