Energy storage lithium battery state of charge assessment method and system integrating physical model and neural network

Abstract

The present invention belongs to the technical field of lithium battery state of charge assessment and discloses a method and system for assessing the state of charge of energy storage lithium batteries that integrates physical models and neural networks. The method comprises: utilizing a lumped parameter equivalent circuit model battery system characterization method to extract charge characteristic information of a power lithium-ion battery; using the charge characteristic information of the power lithium-ion battery as input features of a neural network to construct a novel lumped parameter equivalent circuit model-long short-term memory neural network model; and utilizing the novel lumped parameter equivalent circuit model-long short-term memory neural network model to estimate the state of charge of the power lithium-ion battery to be tested. The present invention provides an innovative and effective solution for SOC estimation of power lithium-ion batteries, which not only simplifies the model construction process but also improves the SOC prediction performance under multiple temperature environments, providing new insights for research and application in related fields.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium battery state of charge assessment, and specifically relates to a method and system for assessing the state of charge of energy storage lithium batteries by integrating a physical model with a neural network. Background Art

[0002] In recent years, with advances in lithium-ion battery technology and the rapid growth of the electric vehicle market, SOC estimation technology for power lithium-ion batteries has become a research focus. Accurate SOC estimation is crucial for optimizing battery performance and ensuring safety. Accurate SOC estimation plays a crucial role in guiding battery management and user behavior. High-precision and robust SOC estimation has become a common goal pursued by researchers and product users. Traditional SOC estimation methods are gradually revealing their limitations, necessitating the introduction of new technologies and methods to improve estimation accuracy and system reliability.

[0003] The accuracy of SOC estimation directly impacts the performance of the battery management system (BMS), and accurate SOC estimation is crucial for electric vehicle range, performance optimization, and battery life. Traditional SOC estimation methods, such as the ampere-hour integration method, the open-circuit voltage method, the electrochemical model method, and equivalent circuit model-based methods, while widely used, are limited in accuracy under complex dynamic conditions and multi-temperature conditions. Equivalent circuit model-based methods describe the electrical characteristics by building an equivalent circuit model of the battery, but require the integration of adaptive optimization filtering algorithms to improve accuracy, which consumes significant computational resources. In recent years, data-driven neural network methods, such as long-short-term memory (LSTM) networks, have become a research hotspot due to their advantages in processing time series data. These methods avoid complex mechanistic analysis and directly use measured voltage and current for SOC prediction. However, unoptimized neural networks suffer from insufficient accuracy in input-output feature maps with strong nonlinear relationships. Researchers have proposed various improved methods, such as LSTM networks combining attention mechanisms with Kalman filtering, bidirectional LSTM networks with integrated attention mechanisms, LSTM estimation networks that incorporate stress, models combining convolutional neural networks (CNNs) and LSTM networks, and hybrid neural network CNN-SAMLSTM models.

[0004] Both traditional estimation methods and data-driven neural network approaches have their own limitations. Traditional methods require the construction of precise equivalent circuit models and the integration of adaptive filtering algorithms to achieve high estimation accuracy. Data-driven methods require not only large amounts of labeled data but also optimization algorithms to adjust network architecture and extract data features. Summary of the Invention

[0005] To solve the problems existing in the prior art, the present invention provides a method and system for evaluating the state of charge of energy storage lithium batteries that integrates physical models and neural networks, providing an innovative and effective solution for SOC estimation of power lithium-ion batteries. It not only simplifies the model building process, but also improves the prediction performance of SOC in multi-temperature environments, providing new ideas for research and application in related fields.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for evaluating the state of charge of an energy storage lithium battery by integrating a physical model with a neural network, the method comprising:

[0008] The battery system characterization method of lumped parameter equivalent circuit model is used to extract the charging characteristic information of power lithium-ion batteries;

[0009] Using the charge characteristic information of the power lithium-ion battery as the input feature of the neural network, a new lumped parameter equivalent circuit model - a long short-term memory neural network model is constructed;

[0010] The new lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested.

[0011] Preferably, the battery system characterization method of the lumped parameter equivalent circuit model is used to extract the charge characteristic information of the power lithium-ion battery, including:

[0012] ;

[0013] Where, is the solution of the quantity to be sought at time k, ρ is the forgetting factor, for +1 time input matrix, K k+1 is the gain at time k+1, P k is the covariance at time k; e k is the prediction error at time k, e max is the maximum value of the error, h is the sensitivity coefficient adjusted by the forgetting factor;

[0014] ;

[0015] Where P is the mean square error; G is the Kalman gain; L is the unit matrix; Q and R are the variances of the state error and measurement error respectively; X k is the state variable [SOC, U P1 , U P2 ] T ;u k is the system input; y k is the observed variable; A kis the state transfer matrix; B k is the system control input matrix; C k 、D k They are the system observation matrix and the driving prediction system observation quantity, U P1 is the polarization voltage, U P2 is the surface effect voltage, is the prior estimate of the state variable at time k+1, To use the state variable X at time k k Calculate the prior estimate of the state variable at time k+1 The calculation formula is: is the posterior estimate of the state variable at time k+1, is the prior estimate of the mean square error at time k+1, is the posterior estimate of the mean square error at time k, is the posterior estimate of the mean square error at time k+1, is the variance of the state error at time k+1, is the Kalman gain at time k+1, is the system observation matrix at time k+1, is the variance of the measurement error at time k+1, is the observation value at time k+1, To use the prior estimate of the state variable at time k+1 Computed state variables.

[0016] Preferably, the charge characteristic information of the power lithium-ion battery is used as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model, including:

[0017] The AFFRLS parameter identification method is used to realize online identification of LPECM parameters, and the EKF algorithm is combined to realize the extraction of charge input characteristics.

[0018] In the basic structure of LSTM, the activation functions sigmoid (σ) and tanh are expressed as:

[0019] ;

[0020] The σ and tanh functions convert the charge input features into nonlinear intervals from 0 to 1 and -1 to 1, respectively. z is the independent variable of the functions σ and tanh.

[0021] The calculation formulas for the three gates in LSTM are:

[0022] ;

[0023] Where i k is the input gate, fk For the forget gate, o k is the output gate, W i , W f and W o is the weight vector of each gate, h k-1 is the output value of the LSTM unit at the previous moment, x k is the input feature at the current moment, b i , b f and b o is the bias term of each gate, k is the time;

[0024] Current cell input information The calculation formula is:

[0025] ;

[0026] Where W c is the weight vector of the current input information, b c is the bias item of the current input information;

[0027] The current input information of the cell is closely related to the cell output value at the previous moment and the input characteristics at the current moment, which contains the new information brought by the current input. The calculation formula of the cell's memory unit and the cell output at the current moment is:

[0028] ;

[0029] Where C k is the cell’s current memory unit, C k-1 is the cell memory unit at the previous moment, h k is the output value of the LSTM unit at the current moment.

[0030] Preferably, the novel lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested, including:

[0031] ;

[0032] Where, is the predicted value of SOC of neural network at time k, represents the input features obtained by feature extraction at time k, is the measured value of the characteristic voltage at time k, Enter the measured value of the characteristic current at time k.

[0033] The present invention also provides a system for evaluating the state of charge of an energy storage lithium battery by integrating a physical model with a neural network. The system is used to implement the aforementioned method. The system includes: an extraction module, a construction module, and an evaluation module;

[0034] The extraction module is used to extract the charge characteristic information of the power lithium-ion battery using a lumped parameter equivalent circuit model battery system characterization method;

[0035] The construction module is used to use the charge characteristic information of the power lithium-ion battery as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model;

[0036] The evaluation module is used to use the novel lumped parameter equivalent circuit model-long short-term memory neural network model to realize the charge state estimation of the power lithium-ion battery to be tested.

[0037] Preferably, the battery system characterization method of the lumped parameter equivalent circuit model is used to extract the charge characteristic information of the power lithium-ion battery, including:

[0038] ;

[0039] Where, is the solution of the quantity to be sought at time k, ρ is the forgetting factor, for +1 time input matrix, K k+1 is the gain at time k+1, P k is the covariance at time k; e k is the prediction error at time k, e max is the maximum value of the error, h is the sensitivity coefficient adjusted by the forgetting factor;

[0040] ;

[0041] Where P is the mean square error; G is the Kalman gain; L is the unit matrix; Q and R are the variances of the state error and measurement error respectively; X k is the state variable [SOC, U P1 , U P2 ] T ;u k is the system input; y k is the observed variable; A k is the state transfer matrix; B k is the system control input matrix; C k 、D k They are the system observation matrix and the driving prediction system observation quantity, U P1 is the polarization voltage, U P2 is the surface effect voltage, is the prior estimate of the state variable at time k+1, To use the state variable X at time k k Calculate the prior estimate of the state variable at time k+1 The calculation formula is: is the posterior estimate of the state variable at time k+1, is the prior estimate of the mean square error at time k+1, is the posterior estimate of the mean square error at time k, is the posterior estimate of the mean square error at time k+1, is the variance of the state error at time k+1, is the Kalman gain at time k+1, is the system observation matrix at time k+1, is the variance of the measurement error at time k+1, is the observation value at time k+1, To use the prior estimate of the state variable at time k+1 Computed state variables.

[0042] Preferably, the charge characteristic information of the power lithium-ion battery is used as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model, including:

[0043] The AFFRLS parameter identification method is used to realize online identification of LPECM parameters, and the EKF algorithm is combined to realize the extraction of charge input characteristics.

[0044] In the basic structure of LSTM, the activation functions sigmoid (σ) and tanh are expressed as:

[0045] ;

[0046] The σ and tanh functions convert the charge input features into nonlinear intervals from 0 to 1 and -1 to 1, respectively. z is the independent variable of the functions σ and tanh.

[0047] The calculation formulas for the three gates in LSTM are:

[0048] ;

[0049] Where i k is the input gate, f k For the forget gate, o k is the output gate, W i , W f and W o is the weight vector of each gate, h k-1 is the output value of the LSTM unit at the previous moment, x k is the input feature at the current moment, b i , b f and b o is the bias term of each gate, k is the time;

[0050] Current cell input information The calculation formula is:

[0051] ;

[0052] Where W c is the weight vector of the current input information, b c is the bias item of the current input information;

[0053] The current input information of the cell is closely related to the cell output value at the previous moment and the input characteristics at the current moment, which contains the new information brought by the current input. The calculation formula of the cell's memory unit and the cell output at the current moment is:

[0054] ;

[0055] Where C k is the cell’s current memory unit, C k-1 is the cell memory unit at the previous moment, h k is the output value of the LSTM unit at the current moment.

[0056] Preferably, the novel lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested, including:

[0057] ;

[0058] Where, is the predicted value of SOC of neural network at time k, represents the input features obtained by feature extraction at time k, is the measured value of the characteristic voltage at time k, Enter the measured value of the characteristic current at time k.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This invention cleverly combines the lumped parameter equivalent circuit model and the LSTM method for estimating the SOC of power lithium batteries, leveraging their respective strengths to form a new LPECM-LSTM fusion estimation strategy. This strategy leverages the advantages of LPECM for preliminary SOC estimation, increases the input features of the LSTM, and reduces the training difficulty of the LSTM. Furthermore, by leveraging the properties of LSTM regression prediction, it further improves the accuracy and robustness of SOC prediction. The contributions of this invention are as follows:

[0061] (1) A lumped parameter equivalent circuit model battery system characterization method is proposed to simulate the charge characteristics of power lithium-ion batteries under complex dynamic electrochemical changes.

[0062] (2) The charge characteristic information of the lumped parameter equivalent circuit model is introduced as the input feature of the neural network, which weakens the nonlinear mapping relationship between the input feature and the output feature, enhances the regression prediction ability of the neural network, and simplifies the neural network hyperparameter adjustment process.

[0063] (3) A new lumped parameter equivalent circuit model - long short-term memory neural network charge estimation strategy was constructed, and the estimation accuracy and robustness of the algorithm were verified. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 Schematic diagram of the LPECM-LSTM estimation algorithm according to an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of a lumped parameter equivalent circuit model according to an embodiment of the present invention;

[0067] Figure 3 Schematic diagram of the SOC estimation results of battery A under DST conditions in an embodiment of the present invention, where (a) and (b) are the SOC estimation results and errors at 15°C, (c) and (d) are the SOC estimation results and errors at 25°C, and (e) and (f) are the SOC estimation results and errors at 35°C;

[0068] Figure 4 Schematic diagram of reliability evaluation of battery A under DST conditions according to an embodiment of the present invention, including (a) ME calculation results of the present invention, (b) RMSE calculation results, (c) MAE calculation results, and (d) MSE calculation results. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0070] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1

[0072] Both traditional estimation methods and data-driven neural network methods have their own limitations. Traditional methods require the construction of precise equivalent circuit models and the integration of adaptive filtering algorithms to achieve high estimation accuracy. Data-driven methods not only require large amounts of labeled data but also employ optimization algorithms to adjust network construction and extract data features. To address these issues, the present invention proposes an innovative hybrid strategy. This strategy combines the advantages of equivalent circuit models and data-driven methods, overcoming some of the shortcomings of existing methods. First, a simple lumped parameter equivalent circuit model is used to perform a preliminary SOC prediction. These preliminary SOC predictions, along with measured voltage and current, are then used as input features to a simple neural network for further SOC prediction. Because the preliminary SOC predictions have a strong linear correlation with the actual SOC, this method can effectively supplement the information in the neural network input, thereby improving the accuracy of SOC prediction. This invention provides an innovative and effective solution for SOC estimation of power lithium-ion batteries. It not only simplifies the model construction process but also improves SOC prediction performance under multiple temperature environments, providing new insights for research and application in related fields.

[0073] This embodiment provides a method for evaluating the state of charge of an energy storage lithium battery by integrating a physical model with a neural network. The method includes:

[0074] The battery system characterization method of lumped parameter equivalent circuit model is used to extract the charging characteristic information of power lithium-ion batteries;

[0075] Using the charge characteristic information of the power lithium-ion battery as the input feature of the neural network, a new lumped parameter equivalent circuit model - a long short-term memory neural network model is constructed;

[0076] The new lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested.

[0077] In this embodiment, a feature extraction method based on LPECM (lumped parameter equivalent circuit model) is used to achieve preliminary extraction of SOC (battery state of charge) features, thereby achieving the purpose of weakening the nonlinear mapping relationship between the input feature parameters and output feature parameters of the neural network. The method of preliminary extraction of SOC feature parameters based on LPECM can fully mine the existing collected information. The extracted SOC features have a strong linear mapping relationship with the predicted values ​​of the neural network, which greatly enhances the convergence of the regression prediction. The constructed LPECM-LSTM algorithm is as follows Figure 1 shown.

[0078] Figure 1The measured current and measured voltage in the figure are commonly used neural network input feature measurements. The present invention uses the LPECM method to extract the charge characteristic parameters based on the measured characteristic parameters. The extraction process is as follows: Figure 1 Nuclear power feature extraction based on extended Kalman filter. In the basic structure of LSTM, the activation functions sigmoid (σ) and tanh are shown in formula (1).

[0079] (1)

[0080] In Equation (1), the σ and tanh functions convert the input features to nonlinear intervals between 0 and 1 and -1 and 1, respectively, enhancing the neural network's ability to extract nonlinear features. z is the independent variable of the σ and tanh functions. By using the concept of gates, LSTM neural networks effectively suppress the vanishing and exploding gradient problems, thereby demonstrating excellent performance in nonlinear regression prediction. The calculation formulas for the three gates in LSTM are shown in Equation (2).

[0081] (2)

[0082] In formula (2), i k is the input gate, f k For the forget gate, o k is the output gate, W i , W f and W o is the weight vector of each gate, h k-1 is the output value of the LSTM unit at the previous moment, x k is the input feature at the current moment, b i , b f and b o is the bias term of each gate, k is the time. The cell input information at the current time The calculation of is shown in formula (3).

[0083] (3)

[0084] In formula (3), W c is the weight vector of the current input information, b c is the bias term for the current input information. The current cell input information is closely related to the cell output value at the previous moment and the input characteristics at the current moment, and contains the new information brought by the current input. The calculation formula for the current cell memory unit and the current cell output is shown in Equation (4).

[0085] (4)

[0086] In formula (4), C k is the cell’s current memory unit, Ck-1 is the cell memory unit at the previous moment, h k is the output value of the LSTM unit at the current moment. Compared with the traditional LSTM algorithm, LPECM-LSTM uses the charge feature extracted by LPECM as an input feature to improve the SOC estimation accuracy. The formula for SOC regression prediction is shown in Equation (5).

[0087] (5)

[0088] In formula (5), is the predicted value of SOC of neural network at time k, represents the input features obtained by feature extraction at time k, is the measured value of the characteristic voltage at time k, Enter the measured value of the characteristic current at time k.

[0089] In this embodiment, a feature extraction method is input based on a lumped parameter equivalent circuit model:

[0090] Accurately characterizing the internal material structure and kinetic process of power lithium-ion batteries during electrochemical reactions is extremely important for SOC feature extraction. The lumped parameter equivalent modeling of power lithium-ion batteries needs to consider the complex electrochemical reaction processes in multiple time scales and multiple spatial dimensions, and then reveal the electrochemical evolution relationship between the lumped parameter equivalent circuit model and the internal materials of the battery to achieve the purpose of accurate characterization. In order to accurately characterize the complex and highly nonlinear reaction mechanism inside power lithium-ion batteries, the present invention studies the internal relationship between the energy conversion process and material evolution inside power lithium-ion batteries, and constructs a lumped parameter equivalent circuit model, such as Figure 2 As shown. oc is the open circuit voltage; U L is the terminal voltage; R1 is the ohmic internal resistance; U R1 Is the ohmic voltage, representing the battery discharge and the battery voltage drop at the moment of discharge end. Represents the RC parallel circuit of the lithium-ion battery polarization effect, consisting of the polarization resistor R P1 and polarization capacitance C P1 Composition, of which U P1 is the polarization voltage; represents the RC parallel circuit of the lithium-ion battery surface effect, consisting of the surface polarization resistance R P2 and surface polarization capacitance C P2 Composition, of which U P2 is the surface effect voltage. The adaptive forgetting factor recursive least squares algorithm is used to realize the online parameter identification of the lumped parameter equivalent circuit model, as shown in Equation (6).

[0091] (6)

[0092] In formula (6), is the solution of the quantity to be sought at time k, ρ is the forgetting factor, for +1 time input matrix, K k+1 is the gain at time k+1, P k is the covariance at time k. k is the prediction error at time k, e max is the maximum value of the error, and h is the sensitivity coefficient adjusted by the forgetting factor.

[0093] The extended Kalman algorithm is used to extract the SOC features of power lithium-ion batteries. The calculation formula is shown in formula (7).

[0094] (7)

[0095] In formula (7), P is the mean square error; G is the Kalman gain; L is the unit matrix; Q and R are the variances of the state error and measurement error, respectively. k is the state variable [SOC, U P1 , U P2 ] T ;u k is the system input; y k is the observed variable; A k is the state transfer matrix, which is used to predict the system variables; B k is the system control input matrix; C k 、D k They are the system observation matrix and the driving prediction system observation quantity, U P1 is the polarization voltage, U P2 is the surface effect voltage, is the prior estimate of the state variable at time k+1, To use the state variable X at time k k Calculate the prior estimate of the state variable at time k+1 The calculation formula is: is the posterior estimate of the state variable at time k+1, is the prior estimate of the mean square error at time k+1, is the posterior estimate of the mean square error at time k, is the posterior estimate of the mean square error at time k+1, is the variance of the state error at time k+1, is the Kalman gain at time k+1, is the system observation matrix at time k+1, is the variance of the measurement error at time k+1, is the observation value at time k+1, To use the prior estimate of the state variable at time k+1 Computed state variables.

[0096] In this embodiment, in order to evaluate the SOC prediction ability of the proposed LPECM-LSTM strategy, the present invention uses four evaluation criteria, namely maximum error (ME), root mean square error (RMSE), mean absolute error (MAE), and mean square error (MSE), as shown in formula (8).

[0097] (8)

[0098] In formula (8), Indicates that the SOC value is predicted using the estimation method at time k, represents the true value of SOC at time k. n represents the total number of prediction steps.

[0099] In this embodiment, the SOC estimation result under the DST condition is:

[0100] from Figure 3 As can be seen from the figure, under DST conditions, the proposed estimation strategy can effectively estimate the SOC of power lithium-ion batteries. The LPECM-LSTM estimation strategy has a significant improvement in prediction accuracy compared to the LSTM estimation method. In order to quantitatively evaluate the effectiveness and performance improvement of the estimation effect in the training set, the four evaluation criteria of maximum error (ME), root mean squared error (RMSE), mean absolute error (MAE), and mean square error (MSE) are used to analyze the prediction effect of the two models in the training set. Figure 4 shown.

[0101] from Figure 4As can be seen, the LECM-LSTM algorithm outperforms both the LSTM and EKF algorithms in terms of ME, RMSE, MAE, and MSE. This demonstrates that the LECM-LSTM algorithm is well-suited to the data from the test set Battery A under DST conditions, achieving regression prediction from input features to SOC. In the Battery A dataset under DST conditions at 15°C, the proposed LECM-LSTM algorithm for power lithium battery SOC estimation achieved ME, RMSE, MAE, and MSE improvements of 34.05%, 35.21%, 31.12%, and 50.02%, respectively, compared to the EKF algorithm, and 81.54%, 57.16%, 48.97%, and 81.65%, respectively, compared to the LSTM algorithm. This was achieved in the Battery A dataset under DST conditions at 25°C. The proposed LECM-LSTM algorithm for power lithium battery SOC estimation achieved improvements in ME, RMSE, MAE, and MSE, respectively, compared to the EKF algorithm by 51.24%, 62.56%, 61.54%, and 85.98%, and compared to the LSTM algorithm by 70.12%, 41.93%, 35.18%, and 66.28%, respectively. In the battery A dataset tested under the 35°C DST operating condition, the proposed LECM-LSTM algorithm achieved improvements in ME, RMSE, MAE, and MSE, respectively, compared to the EKF algorithm by 26.07%, 30.17%, 27.32%, and 51.23%, and compared to the LSTM algorithm by 59.31%, 23.57%, 21.16%, and 41.59%, respectively. It can be seen from the above result graphs and analysis data that in the test set battery A under DST conditions, the estimation accuracy and robustness of the LECM-LSTM algorithm are improved compared with the EKF algorithm and LSTM algorithm.

[0102] This invention cleverly combines the lumped parameter equivalent circuit model and the LSTM method for estimating the SOC of power lithium-ion batteries, leveraging their respective strengths to form a novel LPECM-LSTM fusion estimation strategy. This strategy utilizes LPECM for preliminary SOC estimation while simultaneously incorporating LSTM input features, thereby reducing the difficulty of LSTM training. By fully utilizing existing measurement information, a preliminary method for capturing charge characteristics based on LPECM is proposed. LPECM-extracted features are introduced into LSTM input features, forming a novel LPECM-LSTM regression prediction strategy. Prediction results demonstrate that the proposed LPECM-LSTM fusion estimation strategy exhibits excellent estimation accuracy and robustness, providing a solution for intelligent fusion estimation of the SOC of power lithium-ion batteries.

[0103] Example 2

[0104] The present invention also provides a system for evaluating the state of charge of an energy storage lithium battery by integrating a physical model with a neural network. The system is used to implement the aforementioned method. The system includes: an extraction module, a construction module, and an evaluation module;

[0105] An extraction module is used to extract charge characteristic information of power lithium-ion batteries using a lumped parameter equivalent circuit model battery system characterization method;

[0106] A construction module is used to use the charge characteristic information of the power lithium-ion battery as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model;

[0107] The evaluation module is used to use the novel lumped parameter equivalent circuit model - long short-term memory neural network model to realize the charge state estimation of the power lithium-ion battery to be tested.

[0108] In this embodiment, a lumped parameter equivalent circuit model battery system characterization method is used to extract charging characteristic information of a power lithium-ion battery, including:

[0109] ;

[0110] Where, is the solution of the quantity to be sought at time k, ρ is the forgetting factor, for +1 time input matrix, K k+1 is the gain at time k+1, P k is the covariance at time k; e k is the prediction error at time k, e max is the maximum value of the error, h is the sensitivity coefficient adjusted by the forgetting factor;

[0111] ;

[0112] Where P is the mean square error; G is the Kalman gain; L is the unit matrix; Q and R are the variances of the state error and measurement error respectively; X k is the state variable [SOC, U P1 , U P2 ] T ;u k is the system input; y k is the observed variable; A k is the state transfer matrix; B k is the system control input matrix; C k 、D k They are the system observation matrix and the driving prediction system observation quantity, U P1 is the polarization voltage, U P2 is the surface effect voltage, is the prior estimate of the state variable at time k+1, To use the state variable X at time k k Calculate the prior estimate of the state variable at time k+1 The calculation formula is, is the posterior estimate of the state variable at time k+1, is the prior estimate of the mean square error at time k+1, is the posterior estimate of the mean square error at time k, is the posterior estimate of the mean square error at time k+1, is the variance of the state error at time k+1, is the Kalman gain at time k+1, is the system observation matrix at time k+1, is the variance of the measurement error at time k+1, is the observation value at time k+1, To use the prior estimate of the state variable at time k+1 Computed state variables.

[0113] In this embodiment, the charge characteristic information of the power lithium-ion battery is used as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model, including:

[0114] The AFFRLS parameter identification method is used to realize online identification of LPECM parameters, and the EKF algorithm is combined to realize the extraction of charge input characteristics.

[0115] In the basic structure of LSTM, the activation functions sigmoid (σ) and tanh are expressed as:

[0116] ;

[0117] The σ and tanh functions convert the charge input features into nonlinear intervals from 0 to 1 and -1 to 1, respectively. z is the independent variable of the functions σ and tanh.

[0118] The calculation formulas for the three gates in LSTM are:

[0119] ;

[0120] Where i k is the input gate, f k For the forget gate, o k is the output gate, W i , W f and W o is the weight vector of each gate, h k-1 is the output value of the LSTM unit at the previous moment, x kis the input feature at the current moment, b i , b f and b o is the bias term of each gate, k is the time;

[0121] Current cell input information The calculation formula is:

[0122] ;

[0123] Where W c is the weight vector of the current input information, b c is the bias item of the current input information;

[0124] The current input information of the cell is closely related to the cell output value at the previous moment and the input characteristics at the current moment, which contains the new information brought by the current input. The calculation formula of the cell's memory unit and the cell output at the current moment is:

[0125] ;

[0126] Where C k is the cell’s current memory unit, C k-1 is the cell memory unit at the previous moment, h k is the output value of the LSTM unit at the current moment.

[0127] In this embodiment, the novel lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested, including:

[0128] ;

[0129] Where, is the predicted value of SOC of neural network at time k, represents the input features obtained by feature extraction at time k, is the measured value of the characteristic voltage at time k, Enter the measured value of the characteristic current at time k.

[0130] The present invention proposes an improved lumped parameter equivalent circuit model-long short-term memory neural network strategy for accurate estimation of the SOC of power lithium-ion batteries at multiple temperatures. First, in order to extract the charge characteristic information of power lithium-ion batteries under complex dynamic electrochemical changes, a lumped parameter equivalent circuit model battery system characterization method is proposed. Secondly, in order to weaken the nonlinear mapping relationship between input characteristics and output characteristics, enhance the regression prediction ability of the neural network, and simplify the neural network hyperparameter adjustment process, the charge characteristic information of the lumped parameter equivalent circuit model is introduced as the input feature of the neural network. Finally, a new lumped parameter equivalent circuit model-long short-term memory neural network strategy for power lithium-ion battery SOC prediction is constructed, which uses the lumped parameter equivalent circuit model feature extraction method and the long short-term memory network neural network. The effectiveness and robustness of the proposed strategy are verified under multiple temperatures and working conditions, and the maximum estimation error accuracy can be improved by more than 58.088%. It provides a novel solution for the intelligent fusion prediction strategy of the SOC of power lithium-ion batteries.

[0131] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for evaluating the state of charge of a lithium-ion battery by integrating a physical model with a neural network, characterized in that: The method comprises: The battery system characterization method of the lumped parameter equivalent circuit model is used to extract the charging characteristic information of the power lithium-ion battery; specifically, The adaptive forgetting factor recursive least squares algorithm is used to realize the online parameter identification of the lumped parameter equivalent circuit model. ； Where, is the solution of the quantity to be sought at time k, ρ is the forgetting factor, for +1 time input matrix, K k+1 is the gain at time k+1, P k is the covariance at time k; e k is the prediction error at time k, e max is the maximum value of the error, h is the sensitivity coefficient adjusted by the forgetting factor; The extended Kalman algorithm is used to extract the SOC features of power lithium-ion batteries. ； Where P is the mean square error; G is the Kalman gain; L is the unit matrix; Q and R are the variances of the state error and measurement error respectively; X k is the state variable [SOC, U P1 , U P2 ] T ;u k is the system input; y k is the observed variable; A k is the state transfer matrix; B k is the system control input matrix; C k 、D k They are the system observation matrix and the driving prediction system observation quantity, U P1 is the polarization voltage, U P2 is the surface effect voltage, is the prior estimate of the state variable at time k+1, To use the state variable X at time k k Calculate the prior estimate of the state variable at time k+1 The calculation formula is: is the posterior estimate of the state variable at time k+1, is the prior estimate of the mean square error at time k+1, is the posterior estimate of the mean square error at time k, is the posterior estimate of the mean square error at time k+1, is the variance of the state error at time k+1, is the Kalman gain at time k+1, is the system observation matrix at time k+1, is the variance of the measurement error at time k+1, is the observation value at time k+1, To use the prior estimate of the state variable at time k+1 Calculated state variables; The charge characteristic information of the power lithium-ion battery is used as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model; specifically, the model includes: The AFFRLS parameter identification method is used to realize online identification of LPECM parameters, and the EKF algorithm is combined to realize the extraction of charge input characteristics. In the basic structure of LSTM, the activation functions sigmoid (σ) and tanh are expressed as: ； The σ and tanh functions convert the charge input features into nonlinear intervals from 0 to 1 and -1 to 1, respectively. z is the independent variable of the functions σ and tanh. The calculation formulas for the three gates in LSTM are: ； Where i k is the input gate, f k For the forget gate, o k is the output gate, W i , W f and W o is the weight vector of each gate, h k-1 is the output value of the LSTM unit at the previous moment, x k is the input feature at the current moment, b i , b f and b o is the bias term of each gate, k is the time; Current cell input information The calculation formula is: ； Where W c is the weight vector of the current input information, b c is the bias item of the current input information; The current input information of the cell is closely related to the cell output value at the previous moment and the input characteristics at the current moment, which contains the new information brought by the current input. The calculation formula of the cell's memory unit and the cell output at the current moment is: ； Where C k is the cell’s current memory unit, C k-1 is the cell memory unit at the previous moment, h k is the output value of the LSTM unit at the current moment; The new lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested.

2. The method according to claim 1, characterized in that The novel lumped parameter equivalent circuit model - long short-term memory neural network model is used to estimate the state of charge of the power lithium-ion battery to be tested, including: ； Where, is the predicted value of SOC of neural network at time k, represents the input features obtained by feature extraction at time k, is the measured value of the characteristic voltage at time k, Enter the measured value of the characteristic current at time k.

3. A system for evaluating the state of charge of an energy storage lithium battery that integrates a physical model and a neural network, the system being used to implement the method according to any one of claims 1 to 2, characterized in that: The system includes: an extraction module, a construction module and an evaluation module; The extraction module is used to extract the charge characteristic information of the power lithium-ion battery using a lumped parameter equivalent circuit model battery system characterization method; The construction module is used to use the charge characteristic information of the power lithium-ion battery as the input feature of the neural network to construct a new lumped parameter equivalent circuit model - a long short-term memory neural network model; The evaluation module is used to use the novel lumped parameter equivalent circuit model-long short-term memory neural network model to realize the charge state estimation of the power lithium-ion battery to be tested.

Images