Lithium ion battery charge state estimation method based on parallel battery neural network

Abstract

The invention discloses a lithium ion battery charge state estimation method based on a parallel battery neural network, and the method comprises the steps: obtaining battery charge and discharge data of a parallel battery under different working conditions, carrying out the preprocessing, carrying out the parameter identification through employing a total least square method, carrying out the branch current estimation through employing a DNN deep neural network, and carrying out the calculation of the branch current. And after the estimated current is obtained, SOC prediction is carried out based on a Sage-Husa adaptive filter in combination with an SRCKF (root mean square cubature Kalman filter) algorithm, and the state of charge of the battery is obtained. According to the fusion algorithm, firstly, a first-order ECM is constructed to capture the dynamic response characteristics of each parallel unit, and model parameters are identified by adopting a TLS (Total Least Squares) method, so that parameter deviation caused by measurement noise and system errors is reduced, and the modeling precision is improved. And finally, on the basis of estimating branch current, identifying model parameters and measuring voltage signals, introducing a Sage-Husa adaptive filtering algorithm and combining with a square root cubature Kalman filtering algorithm to realize online estimation of the SOC of each parallel battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium ion battery state of charge estimation, in particular to a lithium ion battery state of charge estimation method based on parallel battery neural network. BACKGROUND

[0002] With global environmental pollution and energy crisis, electric vehicles as a green transportation tool have been rapidly developed. As the core component of electric vehicles, lithium ion batteries have the advantages of high energy density, long cycle life, low self-discharge rate, environmental protection, light weight, etc. Among them, the state of charge (SOC) of the battery is crucial to ensure the safety and efficiency of the battery system. Accurate SOC estimation is not only the key to ensure the safe and efficient operation of the battery system, but also helps to optimize the charging and discharging strategy of the battery, prolong the service life of the battery and improve the overall performance. However, in the assembly process of the battery pack, in order to further improve the capacity and power output, parallel structure is usually introduced. When multiple single batteries are connected in parallel, due to the differences in initial state of charge, internal resistance and aging state between the batteries, there is obvious inconsistency in the branch current between the single batteries. In this way, if the average value of the module current is simply used to replace the actual current of each single battery for SOC estimation, it is easy to cause the accumulation of estimation error, and even cause some batteries to be overcharged or overdischarged, resulting in performance degradation or even safety accidents. Therefore, in the parallel battery pack, accurately estimating the parallel branch current and on this basis estimating the state of charge has become a key problem and research hotspot in parallel battery pack management.

[0003] Therefore, the present application provides a lithium ion battery state of charge estimation method based on parallel battery neural network to solve the problems existing in the prior art, which is a problem that the person skilled in the art urgently needs to solve. SUMMARY

[0004] The purpose of the present application is to provide a lithium ion battery state of charge estimation method based on parallel battery neural network. This "parameter identification-state estimation" mutual coupling, closed-loop updating strategy not only ensures the accuracy of state estimation, but also gives the filter a certain degree of structural adaptive ability.

[0005] To achieve the above purpose, the present application provides the following scheme: A lithium ion battery state of charge estimation method based on parallel battery neural network, comprising the following steps: S1, obtaining the charging and discharging data of lithium ion battery under different working conditions: performing cyclic charging and discharging experiment on lithium ion battery, and recording the time, voltage, current, capacity and temperature data of battery charging and discharging in real time until the discharge capacity is lower than the cut-off voltage of the battery; S2, preprocessing and outlier detection of the time, voltage, current, capacity and temperature obtained in S1; S3, constructing a first-order equivalent circuit model, establishing a state equation and a discretization equation based on battery dynamic characteristics; S4, using total least squares method to identify parameters of the first-order equivalent circuit model, processing errors of input and output data through singular value decomposition, and introducing a sliding window mechanism, rank determination and jump processing, and structure function backstepping model parameters; S5, using deep neural network DNN to estimate parallel branch current, DNN including input layer, hidden layer and output layer, taking total voltage, total current and total battery capacity as input features, and optimizing by minimizing branch current error loss function; S6, based on Sage-Husa adaptive filter combined with root mean square volume Kalman filter algorithm, online estimating battery state of charge, including state prediction, covariance prediction, observation equation calculation, covariance propagation, calculating Kalman gain, state updating and state error calculation, and introducing state protection mechanism to ensure physical rationality of SOC.

[0006] Preferably, in S3, the first-order equivalent circuit model is constructed specifically comprising: The state equation of the first-order equivalent circuit is as follows:

[0007]

[0008] In the formula, U t is the battery terminal output voltage; is the open circuit voltage, which is a function of SOC; U c represents the voltage of the RC network capacitor, which is a state variable; I is the charge and discharge current, and charging and discharging are positive; R 0 , R 1 , C 1 are ohmic resistance, polarization resistance and polarization capacitance respectively; the above equation is discretized to obtain:

[0009]

[0010] In the formula, U c (k) is the RC network capacitor voltage, U c (k +1) is the capacitor voltage at k+1 time, is the sampling interval; (SOC(k)) is the open-circuit voltage at time k. SOC(k) The charging state of the battery at time k; after the first-order equivalent circuit model is established, the model parameters will be identified.

[0011] Preferably, in S4, the total least squares method is used to identify the parameters of the first-order equivalent circuit model, specifically including: S4.1 uses the total least squares method to identify parameters in the first-order equivalent circuit model. The core idea is to minimize the perpendicular distance from the data points to the fitted model, thereby handling the situation where both input and output data contain noise. Let the fitted model be... After considering observation errors, it is expressed as The goal of total least squares is to minimize the error, i.e.:

[0012]

[0013] in, for Norms, where X and y are noisy data matrices.

[0014] S4.2. At each step, a small sliding window is used for local identification; the constructed observation matrix is ​​processed through singular value decomposition to obtain a stable minimum norm solution: singular value decomposition is performed on the target matrix, and then the right singular value vector with the minimum singular value is obtained, thereby solving the... First, construct an extended observation matrix, perform singular value decomposition on A, and take the right singular vector v corresponding to the smallest singular value. The specific formula is as follows:

[0015]

[0016]

[0017]

[0018] S4.3, Combining the discretized state-space equations of the first-order equivalent circuit model, we obtain:

[0019] in, Given terms, ,in, Open circuit voltage, Given the input vector, the expression is: , The internal resistance of the battery is ohmic. This is the battery's operating current. and They are respectively and , The parameters to be identified;

[0020]

[0021]

[0022] in, It is a time constant. C 1 It is a polarizing capacitor.

[0023] Preferably, in S5, a deep neural network (DNN) is used to estimate the current in the parallel branches, specifically including: S5.1. Given the total current and the current of one branch, a deep neural network (DNN) is a combination of multiple nonlinear mapping functions, in the following form:

[0024] Where X is the input feature vector. This is a predicted value; S5.2. A deep neural network consists of the following structure: an input layer, a hidden layer, and an output layer, wherein the hidden layer consists of multiple layers;

[0025] in, For the input vector, The total voltage is the current. d represents the total current, Cap represents the total battery capacity, and d represents the dimension. Let be a d-dimensional real space; The hidden layer is:

[0026]

[0027]

[0028] in, It is an activation function. For the first i The weight matrix of the layer, For the first i Layer bias vector, Let i be the output vector of the i-th hidden layer of the DNN neural network; S5.3 The optimization objective of a deep neural network is to minimize the loss function. The expression for the loss function, using branch current error as the loss function, is as follows:

[0029] in, Here is the defined loss function, and N is the number of samples. and Let be the model's predicted values ​​for the currents in two branches of i samples.

[0030] Preferably, in S6, the battery state of charge is estimated online based on the Sage-Husa adaptive filter combined with the root mean square capacitive Kalman filter algorithm, specifically including: (1) State prediction: The total least squares method is used to identify parameters and decouple the filter. At each time step, the total least squares method is used to estimate the parameters of the first-order equivalent circuit model in real time, and the system matrix is ​​dynamically constructed.

[0031] in, Let A be the state vector, A be the state transition matrix, and B be the input matrix. This is the battery's nominal capacity; (2) Covariance prediction:

[0032] Where Q is the process noise covariance matrix. , These are the predicted covariance matrix and the covariance matrix of the previous time step, respectively. (3) Introduce a state protection mechanism: if the voltage disturbance state If the absolute value is less than 1e-5, the value from the previous time step is directly inherited; if the SOC tends to be less than 0 or greater than 1, then a limit is applied. (4) Covariance propagation uses square root decomposition, while the root mean square volume Kalman filter algorithm propagates the covariance by performing Cholesky decomposition on the covariance matrix, i.e.:

[0033] Then construct the cubature sigma point:

[0034] in, Let be the set of volume points sigma points. Estimate the prior state at the current moment. The scaling parameters for the volumetric Kalman filter. It is the square root of the prior covariance; (5) Calculation of observation equations:

[0035] in, This is the first state vector of the model, representing the voltage of the RC circuit. This is the second state vector of the model, denoted as SOC; (6) Calculate the Kalman gain:

[0036] (7) Status update:

[0037] in, This is the posterior state estimate. These are the prior state estimates. These are the actual observed values. The predicted value is the observed value. Calculate the observed values The truth value of; (8) Calculation of state error:

[0038] in, For state error, This is the actual state vector at the current moment. Here is the state transition matrix. The state vector of the previous time step, For the input matrix, This is the system's input vector.

[0039] Preferably, S6 also includes: Within the framework of integrating the Sage-Husa adaptive filter and the root mean square occult Kalman filter algorithm, the Sage-Husa adaptive filter analyzes the noise covariance matrix Q and the observation noise covariance matrix R of the system residual dynamic adjustment process in real time, and feeds the updated noise statistics back to the root mean square occult Kalman filter algorithm. At the same time, the root mean square occult Kalman filter algorithm uses these adaptively estimated noise parameters to perform high-precision state prediction and update calculations, and its output state estimation residuals are used as inputs to the Sage-Husa algorithm to form a closed-loop optimization. Specifically, the core mechanism of the Sage-Husa adaptive filter lies in: based on the maximum a posteriori estimation criterion, it achieves online recursive estimation of the noise covariance matrix through the statistical properties of the information sequence within the sliding window; this adaptive idea is innovatively introduced into the root mean square occlusive Kalman filter architecture, and real-time estimation models of the following forms are established for process noise and observation noise respectively: Process noise Q update:

[0040]

[0041] in, and This is the forgetting factor, typically taken as 0.01 to 0.1. Let be the process noise covariance matrix at time k. Let be the state error vector. Let be the observation noise covariance at time k. This represents the actual observed value at time k.

[0042] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a lithium-ion battery state-of-charge estimation method based on a parallel battery neural network as described above.

[0043] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: Based on the aforementioned parameter identification, current estimation, and SOC estimation steps, the method of this invention can update model parameters online, including ohmic resistance, polarization resistance, polarization capacitance, and SOC. The proposed Sage-Husa adaptive filter combined with the SRCKF algorithm can accurately estimate the battery's SOC. This strategy of coupling parameter identification and state estimation in a closed-loop update ensures the accuracy of state estimation while also endowing the filter with a certain degree of structural adaptability, making it particularly suitable for dynamic modeling of batteries as they age over time and experience fluctuations in charge / discharge rates. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating the lithium-ion battery state-of-charge estimation method based on parallel battery neural networks provided by this invention. Figure 2 This is a schematic diagram comparing the actual state of charge (SOC) value with the estimated SOC value in an embodiment of the present invention. Figure 3 This is a diagram showing the state-of-charge error curve in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the actual current and the estimated current in an embodiment of the invention; Figure 5 This is a current error curve diagram in an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1 like Figure 1 As shown, the present invention provides a method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network, comprising the following steps: S1. Obtain data on lithium-ion batteries under different operating conditions: Conduct cyclic charge-discharge experiments on lithium-ion batteries under different operating conditions until the discharge capacity is lower than the battery's cutoff voltage, and record data such as battery charging and discharging time, voltage, current, capacity, and temperature in real time.

[0049] S2. Data processing: Preprocess the time, voltage, current, capacity and temperature obtained in step 1, and detect outliers.

[0050] S3. Constructing a first-order equivalent circuit model: The MCE structure is simple, computationally efficient, and suitable for real-time battery management. The specific construction process is as follows: S3.1 The state equations of the first-order equivalent circuit are as follows:

[0051]

[0052] In the formula, U t This is the output voltage of the battery terminals; This is the open-circuit voltage, a function of SOC; U c This represents the voltage across the capacitor in the RC network; this voltage is a state variable. I This is the charging and discharging current; charging and discharging are positive. R 0 , R 1 , C 1 Let be the ohmic internal resistance, polarization internal resistance, and polarization capacitance, respectively; discretizing the above equations yields:

[0053]

[0054] In the formula, U c (k) represents the capacitor voltage of the RC network, which is k; U c (k+1) represents the capacitor voltage at time k+1, and is the sampling interval; (SOC(k)) represents the open-circuit voltage (OCV) at time k, which is a function of SOC; SOC(k) represents the state of charge of the battery at time k. After the model is established, the model parameters will be identified.

[0055] S4. TLS Parameter Identification: The Total Least Squares (TLS) method can address the error problem of coexisting input and output data through singular value decomposition, effectively reducing parameter identification errors and correcting errors in the regression matrix and observation vector. The specific steps are as follows: S4.1 This invention employs the TLS method to identify parameters in the equivalent circuit model. Its core idea is to minimize the perpendicular distance from data points to the fitted model, thereby handling situations where both input and output data contain noise. Unlike ordinary least squares (OLS), which only considers output error, TLS maintains high identification accuracy even with input error. Let the fitted model be... After considering observation errors, it can be expressed as ; The goal of TLS is to minimize errors, i.e.

[0056]

[0057] in, for Norms, where X and y are noisy data matrices.

[0058] S4.2. Local identification is performed at each step using only a small sliding window (within 100 points), avoiding fitting the entire sequence and dynamically tracking the time-varying nature of battery parameters. The constructed observation matrix is ​​processed using Singular Value Decomposition (SVD) to obtain a stable minimum norm solution: Singular value decomposition is performed on the target matrix, and then the right singular value vector with the minimum singular value is obtained, thus yielding the... To determine the value, first construct an extended observation matrix, perform singular value decomposition on A, and then take the right singular vector v corresponding to the minimum singular value. The specific formula is as follows:

[0059]

[0060]

[0061]

[0062] S4.3. With the core idea of ​​TLS, the state-space equations that can be discretized from the equivalent circuit model are:

[0063] Then, rearranging into linear form, we have:

[0064] Simplified to:

[0065] S4.4, The column vector corresponding to the minimum singular value of V for the TLS solution. After normalization, we get Where n is the number of parameters. Parameter extraction, from... Inverse model parameters:

[0066]

[0067]

[0068] in, It is a time constant. R 1 Polarization resistor, C 1 It is a polarizing capacitor.

[0069] S4.5 Three improvements to the TLS identification process in this invention: S4.5.1. Sliding Window TLS Identification Mechanism: This invention introduces a 100-point sliding window to perform TLS identification locally in real time, enhancing the model's ability to track the time-varying nature of battery parameters and avoiding prediction lag caused by global static fitting.

[0070] S4.5.2. Introduction of Rank Detection and Skip-Step Handling: A matrix rank detection mechanism is added before TLS solution. If the observed matrix A is not full rank, the current update is skipped to avoid parameter mutation or irreversible solution failure.

[0071] S4.5.3. Inverse model parameter derivation based on structure function: TLS only identifies intermediate variables, and the final parameters such as R1 and C1 are obtained through inverse structural function derivation, thereby enhancing the physical consistency of the model and avoiding unreasonable predictions caused by blind mathematical fitting.

[0072] S5. Current Estimation: A deep neural network (DNN) is used to predict the current in the two branches. The data used is the raw data from S1, which comes from the charging and discharging data of parallel batteries in the laboratory under various operating conditions, including voltage, total current, branch current, capacity, and temperature. The following is the DNN implementation process. S5.1 In parallel battery packs, independent current estimation for each battery can significantly improve the accuracy of SOC estimation because differences in battery parameters lead to uneven current distribution, and directly measuring or estimating the current of a single battery can avoid errors in current sharing assumptions. However, independent current estimation helps identify inconsistencies between batteries, such as aging differences, and enhances fault diagnosis capabilities. This is especially relevant when the total current and the current of one branch are known. Deep neural networks (DNNs) are combinations of multiple layers of nonlinear mapping functions, in the following form:

[0073] Where X is the input feature vector, for example , These are predicted values.

[0074] S5.2. Deep neural networks generally consist of an input layer, hidden layers, and an output layer, with the hidden layer composed of multiple layers.

[0075]

[0076] in, For the input vector, The total voltage is the current. d represents the total current, Cap represents the total battery capacity, and d represents the dimension. Let be a d-dimensional real space; The hidden layer is:

[0077]

[0078]

[0079] in, It is an activation function, such as : , Let be the weight matrix of the l-th layer. is the bias vector of the l-th layer. Let i be the output vector of the i-th hidden layer of the DNN neural network; S5.3 The optimization objective of a deep neural network is to minimize the loss function. Using the branch current error as the loss function, we have the following expression:

[0080] in, Here is the defined loss function, and N is the number of samples. and Let be the model's predicted values ​​for the currents in two branches of i samples.

[0081] like Figure 2 As shown, the solid line represents the actual value of the branch current, and the dashed line represents the predicted value from the DNN neural network. It is easy to see that the actual and predicted values ​​almost coincide. This preliminarily verifies the accuracy and reliability of the algorithm. Figure 3 This is the error between the predicted current and the actual current, which can be seen from the graph as being less than 0.05A, which is relatively small.

[0082] S6. SOC Estimation Algorithm: Based on the previous parameter identification and current estimation results, a Suge-Husa filter fused with the Root Mean Square Capacitive Kalman Filter (SRCKF) algorithm is used to address the complexity of the BMS system. In traditional Kalman filtering algorithms, the process noise covariance matrix Q and the observation noise covariance matrix R are usually set to fixed values ​​before filtering begins. However, during actual battery operation, the system noise exhibits strong non-stationarity and uncertainty due to factors such as ambient temperature, operating current, load disturbance, and aging degree. Against this backdrop, the Sage-Husa adaptive noise estimation method is introduced into the SRCKF algorithm to construct a robust filtering algorithm with dynamic adaptability, improving the overall accuracy and stability of state estimation and voltage prediction. The following are the key improvements and innovative details of this method's fusion process. The specific calculation process is as follows: S6.1 Based on the battery model, the discretized state equation and observation equation can be obtained:

[0083]

[0084] in, The voltage across the RC network capacitor. Sampling time; This refers to the battery's nominal capacity. This is a polynomial fit for the open-circuit voltage.

[0085] S6.2 and SRCKF are improved versions of the unscented Kalman filter (UKF), employing square root decomposition to enhance numerical stability. Their core steps include prediction and updating. Step 1: State Prediction, TLS Parameter Identification, and Filter Decoupling to Improve Model Dynamic Accuracy. To ensure the structural accuracy of the filter state-space modeling, the total least squares (TLS) method is used at each time step to estimate the first-order RC equivalent circuit parameters in real time and dynamically construct the system matrix.

[0086] Where: X is the state vector. A is the state transition matrix. :describe Dynamic decay; :describe Dynamic decay; : Integral properties of SOC; B is the input matrix, , Input current I pair The impact; The cumulative effect of input current I on SOC; This is the battery's nominal capacity; 3600 is used for unit conversion, from seconds to hours.

[0087] Step 2: Covariance Prediction

[0088] in, Q Let L be the process noise covariance matrix, and L be a lower triangular matrix. , These are the predicted covariance matrix and the covariance matrix of the previous time step, respectively. Step 3: The state variable protection mechanism ensures the rationality and physical interpretability of the filtering. In actual estimation, due to drastic noise fluctuations or imperfect models, the filtering update results may exhibit non-physical or unreasonable states, such as the voltage residual becoming negative or the SOC exceeding the [0,1] range. To address this, the present invention introduces a state protection mechanism: if the absolute value of the voltage disturbance state Vc is too small, less than 1e-5, it may be caused by numerical instability, and the value from the previous moment is directly inherited; if the SOC tends to be less than 0 or greater than 1, then restrictions are imposed.

[0089] Step 4: Covariance propagation employs square root decomposition to enhance numerical stability and computational efficiency. Compared to the standard CKF, SRCKF propagates the covariance by performing Cholesky decomposition on the covariance matrix. This method effectively avoids covariance matrix divergence caused by rounding errors or non-positive definiteness, while reducing computational complexity and improving the numerical stability of the filter in embedded systems or long-term operation. It forms the structural basis for combining Sage-Husa and SRCKF. That is:

[0090] Then construct the cubature sigma point:

[0091] in, Let be the set of volume points sigma points. Estimate the prior state at the current moment. The scaling parameters for the volumetric Kalman filter. It is the square root of the prior covariance; Step 5: Calculation of observation equations

[0092] in, Calculate the observed values The truth value of , This is the first state vector of the model, representing the voltage of the RC circuit. This is the second state vector of the model, denoted as SOC; Step 6: Calculate the Kalman gain

[0093] in, For Kalman filter gain, The state-observation cross covariance matrix, To observe the inverse of the covariance matrix; Step 7: State Update

[0094] in: : Posterior state estimate Prior state estimate (predicted value). : The actual observed value.

[0095] Step 8: Calculation of State Error

[0096] in: State error This is the actual state vector at the current moment. Here is the state transition matrix. The state vector of the previous time step, For the input matrix, The input vector of the system; S6.3, A collaborative estimation method integrating Sage-Husa adaptive filtering and root mean square volumetric Kalman filtering (SRCKF). In this framework, the Sage-Husa algorithm dynamically adjusts the process noise covariance matrix Q and the observation noise covariance matrix R by analyzing the system residuals in real time, and feeds the updated noise statistics back to the SRCKF algorithm. Simultaneously, SRCKF uses these adaptively estimated noise parameters to perform high-precision state prediction and update calculations, and its output state estimation residuals serve as input to the Sage-Husa algorithm to form a closed-loop optimization. Specifically, the core mechanism of the Sage-Husa filter lies in: based on the maximum a posteriori estimation criterion, it achieves online recursive estimation of the noise covariance matrix through the statistical characteristics of the information sequence within the sliding window. This invention innovatively introduces this adaptive idea into the SRCKF architecture, establishing real-time estimation models of the following forms for process noise and observation noise respectively: Process noise Q update:

[0097]

[0098] in: This is the forgetting factor (usually taken as 0.01-0.1). This is the forgetting factor (usually taken as 0.01-0.1). Figure 4 This is a comparison chart of the actual and predicted SOC values ​​for two branches. The solid line represents the actual value, and the dashed line represents the predicted value. Figure 5 It is the error between the predicted and actual values ​​of the two branches, which can be obtained from... Figure 4 It can be seen that the predicted values ​​of branch 1 and branch 2 basically coincide with the actual values. Figure 5 It can also be seen that the error between the two branches is very small, within 0.02, indicating the accuracy and reliability of the algorithm.

[0099] Based on the parameter identification process, current estimation process, and SOC estimation steps described above, model parameters, including ohmic resistance, polarization resistance, polarization capacitance, and SOC, can be updated online. The proposed Sage-Husa adaptive filter combined with the SRCKF algorithm can accurately estimate the battery's SOC.

[0100] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a lithium-ion battery state-of-charge estimation method based on a parallel battery neural network as described above.

[0101] Specifically, as shown in Table 1 (Statistical Error of State of Charge (SOC) Estimation): Table 1

[0102] The effectiveness of the proposed method for estimating the state of charge (SOC) of parallel battery packs was verified through extensive testing under four typical operating conditions: constant current constant voltage (CCCV) charging, Urban Energy Consumption Standard (UDDS), Federal Urban Energy Consumption Standard (FUDS), and a hybrid drive cycle (DST+UDDS+FUDS). The estimation error remained within 3%. The error statistics for each cycle are summarized in the table. The CCCV charging cycle performed best due to its relatively stable operation and gradual changes in current and voltage. For branch 2, the mean absolute error (MAE) and root mean square error (RMSE) of the estimated SOC were as low as 0.08% and 0.67%, respectively. The error was slightly higher in the hybrid drive cycle, with the MAE and RMSE for branch 1 reaching 1.35% and 1.42%, respectively. This is mainly because the hybrid cycle involves multiple dynamic cycles with rapid changes in current and voltage. However, the estimation accuracy still meets the requirements of practical engineering applications. The proposed state of charge (SOC) estimation method is evaluated using various error metrics, including mean absolute error (MAE), maximum absolute error (MAXE), mean square error (MSE), and root mean square error (RMSE), demonstrating its adaptability and reliability across different driving cycles and providing an effective solution for practical battery management system (BMS) applications.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0104] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for estimating the state of charge (SOC) of a lithium-ion battery based on a parallel battery neural network, characterized in that, Includes the following steps: S1. Obtain charging and discharging data of lithium-ion batteries under different operating conditions: Conduct cyclic charging and discharging experiments on lithium-ion batteries, and record the charging and discharging time, voltage, current, capacity and temperature data of the batteries in real time until the discharge capacity is lower than the battery's cutoff voltage. S2. Preprocess and detect outliers on the time, voltage, current, capacity and temperature obtained in S1; S3. Construct a first-order equivalent circuit model and establish state equations and discretized equations based on the dynamic characteristics of the battery. S4. The total least squares method is used to identify the parameters of the first-order equivalent circuit model. The error of the input and output data is processed by singular value decomposition. The sliding window mechanism, rank judgment and step skipping are introduced, and the model parameters are back-inferred by the structure function. S5. A deep neural network (DNN) is used to estimate the current in parallel branches. The DNN includes an input layer, a hidden layer, and an output layer. The total voltage, total current, and total battery capacity are used as input features, and optimization is performed by minimizing the branch current error loss function. S6. Based on the Sage-Husa adaptive filter combined with the root mean square capacitive Kalman filter algorithm, the battery state of charge is estimated online, including state prediction, covariance prediction, observation equation calculation, covariance propagation, Kalman gain calculation, state update and state error calculation, and a state protection mechanism is introduced to ensure the physical rationality of the SOC.

2. The method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network according to claim 1, characterized in that, In step S3, constructing the first-order equivalent circuit model specifically includes: The state equations of the first-order equivalent circuit are as follows: In the formula, U t This is the output voltage of the battery terminals; This is the open-circuit voltage, a function of SOC; U c This represents the voltage across the capacitor in the RC network; this voltage is a state variable. I This is the charging and discharging current; charging and discharging are positive. R 0 , R 1 , C 1 Let be the ohmic internal resistance, polarization internal resistance, and polarization capacitance, respectively; discretizing the above equations yields: In the formula, U c (k) represents the capacitor voltage of the RC network. U c (k+1) is the capacitor voltage at time k+1. The sampling interval; (SOC(k)) is the open-circuit voltage at time k. SOC(k) The charging state of the battery at time k; after the first-order equivalent circuit model is established, the model parameters will be identified.

3. The method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network according to claim 1, characterized in that, In step S4, the total least squares method is used to identify the parameters of the first-order equivalent circuit model, specifically including: S4.1 uses the total least squares method to identify parameters in the first-order equivalent circuit model. The core idea is to minimize the perpendicular distance from the data points to the fitted model, thereby handling the situation where both input and output data contain noise. Let the fitted model be... After considering observation errors, it is expressed as The goal of total least squares is to minimize the error, i.e.: in, for Norm, X and y are noisy data matrices. S4.2, Local identification is performed using a small sliding window at each step; the constructed observation matrix is ​​processed by singular value decomposition to obtain a stable minimum norm solution: singular value decomposition is performed on the target matrix, and then the right singular value vector with the minimum singular value is obtained, thereby obtaining the norm. First, construct an extended observation matrix, perform singular value decomposition on A, and take the right singular vector v corresponding to the smallest singular value. The specific formula is as follows: S4.3, Combining the discretized state-space equations of the first-order equivalent circuit model, we obtain: in, Given terms, ,in, Open circuit voltage, Given an input vector, the expression is: , The ohmic internal resistance of the battery. This is the battery's operating current. and They are respectively and , The parameters to be identified; S4.4, The column vector corresponding to the smallest singular value of V in the total least squares solution. After normalization, we get ,in, n Number of parameters; Parameter extraction: from Inverse model parameters: in, It is a time constant. C 1 It is a polarizing capacitor.

4. The method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network according to claim 1, characterized in that, In step S5, a deep neural network (DNN) is used to estimate the current in the parallel branches, specifically including: S5.

1. Given the total current and the current of one branch, a deep neural network (DNN) is a combination of multiple nonlinear mapping functions, in the following form: Where X is the input feature vector. This is a predicted value; S5.

2. A deep neural network consists of the following structure: an input layer, a hidden layer, and an output layer, wherein the hidden layer consists of multiple layers; in, For the input vector, The total voltage is the current. d represents the total current, Cap represents the total battery capacity, and d represents the dimension. Let be a d-dimensional real space; The hidden layer is: in, It is an activation function. For the first i The weight matrix of the layer, For the first i Layer bias vector, Let i be the output vector of the i-th hidden layer of the DNN neural network; S5.3 The optimization objective of a deep neural network is to minimize the loss function. The expression for the loss function, using branch current error as the loss function, is as follows: in, Here is the defined loss function, and N is the number of samples. and Let be the model's predicted values ​​for the currents in two branches of i samples.

5. The method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network according to claim 1, characterized in that, In step S6, the battery state of charge is estimated online based on the Sage-Husa adaptive filter combined with the root mean square capacitive Kalman filter algorithm, specifically including: (1) State prediction: The total least squares method is used to identify parameters and decouple and couple them with the filter. At each time step, the total least squares method is used to estimate the parameters of the first-order equivalent circuit model in real time, and the system matrix is ​​dynamically constructed. in, Let A be the state vector, A be the state transition matrix, and B be the input matrix. This is the battery's nominal capacity; (2) Covariance prediction: Where Q is the process noise covariance matrix. , These are the predicted covariance matrix and the covariance matrix of the previous time step, respectively. (3) Introduce a state protection mechanism: if the voltage disturbance state If the absolute value is less than 1e-5, the value from the previous time step is directly inherited; if the SOC approaches 0 or is greater than 1, then a limit is applied. (4) Covariance propagation uses square root decomposition, while the root mean square volume Kalman filter algorithm propagates the covariance by performing Cholesky decomposition on the covariance matrix, i.e.: Then construct the cubature sigma point: in, Let be the set of volume points sigma points. Estimate the prior state at the current moment. The scaling parameters for the volumetric Kalman filter. It is the square root of the prior covariance; (5) Calculation of observation equations: in, This is the first state vector of the model, representing the voltage of the RC circuit. This is the second state vector of the model, denoted as SOC; (6) Calculate the Kalman gain: in, For Kalman filter gain, The state-observation cross covariance matrix, The inverse of the observation covariance matrix; (7) Status update: in, This is the posterior state estimate. These are the prior state estimates. These are the actual observed values. The predicted value is the observed value. Calculate the observed values The truth value of; (8) Calculation of state error: in, For state error, This is the actual state vector at the current moment. Here is the state transition matrix. The state vector of the previous time step, For the input matrix, This is the system's input vector.

6. The method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network according to claim 5, characterized in that, S6 also includes: Within the framework of integrating the Sage-Husa adaptive filter and the root mean square volumetric Kalman filter algorithm, the Sage-Husa adaptive filter analyzes the noise covariance matrix Q and the observation noise covariance matrix R of the system residual dynamic adjustment process in real time, and feeds the updated noise statistics back to the root mean square volumetric Kalman filter algorithm. At the same time, the root mean square volumetric Kalman filter algorithm uses these adaptively estimated noise parameters to perform high-precision state prediction and update calculations, and its output state estimation residuals are used as inputs to the Sage-Husa algorithm to form a closed-loop optimization. Specifically, the core mechanism of the Sage-Husa adaptive filter lies in: based on the maximum a posteriori estimation criterion, it achieves online recursive estimation of the noise covariance matrix through the statistical characteristics of the time series within the sliding window; this adaptive idea is innovatively introduced into the root mean square occlusive Kalman filter architecture, and real-time estimation models of the following forms are established for process noise and observation noise respectively: Process noise Q update: in, and This is the forgetting factor, typically taken as 0.01 to 0.

1. Let be the process noise covariance matrix at time k. Let be the state error vector. Let be the observation noise covariance at time k. This represents the actual observed value at time k.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for estimating the state of charge of a lithium-ion battery based on a parallel battery neural network as described in any one of claims 1 to 6.

Images