Method and device for estimating state of charge of battery

Abstract

The embodiment of the invention provides a battery charge state estimation method and device, and the method comprises the steps: building a Hammerstein battery model based on an extreme learning machine, and enabling the Hammerstein battery model to comprise an extreme learning machine nonlinear block and a dynamic linear block which are connected in series; separating parameters of a nonlinear block and parameters of a dynamic linear block in the Hammerstein battery model, identifying the parameters of the dynamic linear block by using a maximum likelihood stochastic gradient algorithm, and identifying an output weight of the nonlinear block of an extreme learning machine by using a least square method; the terminal voltage and the terminal current of the battery are input into the Hammerstein battery model to estimate the state of charge of the battery, so that the estimation precision of the state of charge of the battery is improved, the calculation complexity is reduced, and the application range is wider.

Description

Technical Field

[0001] The present invention relates to the field of batteries, and in particular, to a method and device for estimating the state of charge of a battery. Background Art

[0002] The state of charge (SOC) of a battery is an important basis for judging whether the battery is overcharged or over-discharged and ensuring the reliable application of the battery. However, the SOC represents an internal state quantity of the battery rather than a specific physical quantity and cannot be directly measured.

[0003] Based on traditional experimental testing methods, SOC estimation is obtained by using an estimation equation between the SOC and battery measurement variables. There are mainly three methods: the ampere-hour integration (AHI) method, the open-circuit voltage (OCV) method, and the electrochemical impedance spectroscopy (EIS) method. The ampere-hour integration (AHI) method obtains a real-time state of charge (SOC) estimation by integrating the measured charge / discharge current of the battery. However, the initial state of charge (SOC) and real-time health status of the battery affect the accuracy of the SOC estimation. In the open-circuit voltage (OCV) method, the estimation of the battery state of charge (SOC) is determined by the position of the measured open-circuit voltage (OCV) on the OCV-SOC characteristic curve, which is obtained through offline sampling in the laboratory. The electrochemical impedance spectroscopy (EIS) method is difficult to achieve online estimation. Therefore, it is necessary to propose a new method for estimating the state of charge of a battery. Summary of the Invention

[0004] In view of the above problems, a method and device for estimating the state of charge of a battery are proposed to overcome the above problems or at least partially solve the above problems, including:

[0005] Construct a Hammerstein battery model based on an extreme learning machine, where the Hammerstein battery model includes a series connection of an extreme learning machine non-linear block and a dynamic linear block;

[0006] Separate the parameters of the non-linear block in the Hammerstein battery model from the parameters of the dynamic linear block, identify the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, and identify the output weights of the extreme learning machine non-linear block using the least squares method;

[0007] Input the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the state of charge of the battery.

[0008] Optionally, the Hammerstein model constructed based on the extreme learning machine is:

[0009] A(q)SOC(k) = B(q)s(k) + D(q)v(k)

[0010]

[0011] where U(k) = [u 1 (k), u 2 (k)] T = [V(k), I(k)] T ∈ R 2 is composed of the battery terminal voltage and terminal current. The SOC of the battery is the output SOC(k). D(q)v(k) is the output of an MA model driven by white noise v(k). The linear module is a controlled autoregressive moving average (CARMA) model. s(k) is the extreme learning machine neural network nonlinear block. W o i is the weight of the output layer of the extreme learning machine. W h ji (j = 1, 2, i = 1, 2,..., l) are the weights of the hidden layer of the extreme learning machine. b i is the threshold of the i-th hidden node. g(*) is the activation function. A(q), B(q) and D(q) are polynomials of known order with the backward shift operator q -1 defined as:

[0012] A(q): = 1 + a 1 q -1 +... + a n q -n ,

[0013] B(q): = b 0 + b 1 q -1 +... + b n q -n ,

[0014] D(q): = 1 + d 1 q -1 +... + d n q -n .

[0015] Optionally, the matrix form of the extreme learning machine nonlinear module s(k) is as follows:

[0016] H = g[W h U(k) + b],

[0017] s(k) = H T W o ,

[0018] where the activation function g(*) and the input vector U(k) are as follows:

[0019]

[0020] Optionally, the linear module of the Hammerstein battery model is as follows:

[0021]

[0022] Optionally, separating the parameters of the non-linear block in the Hammerstein battery model from the parameters of the dynamic linear block, including:

[0023] Separating the parameter W of the extreme learning machine non-linear block through the key-term separation technique o from the parameter θ = [a 1 ,...a n ,b 1 ,...b n ,d 1 ,...d n of the dynamic linear block T wherein b i is decoupled.

[0024] Optionally, identifying the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, including:

[0025] For the parameters θ of the dynamic linear block, iteratively update using the following maximum likelihood stochastic gradient algorithm:

[0026]

[0027] where is the information vector containing historical SOC, historical non-linear block output, and noise, and γ(k) is the iteration step size.

[0028] Optionally, identifying the output weights of the extreme learning machine non-linear block using the least squares method, including:

[0029] For the output weights W o of the extreme learning machine non-linear block, calculate using the following least squares method:

[0030]

[0031] where H is the hidden layer output matrix, and H + is its Moore-Penrose generalized inverse.

[0032] Optionally, before inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model, it further includes:

[0033] Normalizing the terminal voltage and the terminal current.

[0034] Optionally, the method further includes;

[0035] Divide data segments according to real-time battery operating conditions parameters, and independently optimize normalization parameters for each segment of data. Among them, the operating conditions parameters include at least one of the temperature change rate, charge-discharge switching state, and load fluctuation intensity.

[0036] A state of charge estimation device for a battery, comprising:

[0037] A construction module, configured to construct a Hammerstein battery model based on an extreme learning machine. The Hammerstein battery model includes a series connection of an extreme learning machine non-linear block and a dynamic linear block;

[0038] A parameter processing module, configured to separate the parameters of the non-linear block in the Hammerstein battery model from the parameters of the dynamic linear block, identify the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, and identify the output weights of the extreme learning machine non-linear block using the least squares method;

[0039] An estimation module, configured to input the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the state of charge of the battery.

[0040] In the embodiment of the present invention, by constructing a Hammerstein battery model based on an extreme learning machine, the Hammerstein battery model includes a series connection of an extreme learning machine non-linear block and a dynamic linear block; separating the parameters of the non-linear block in the Hammerstein battery model from the parameters of the dynamic linear block, and identifying the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, and identifying the output weights of the extreme learning machine non-linear block using the least squares method; inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the state of charge of the battery, the estimation accuracy of the state of charge of the battery is improved, and at the same time, the calculation complexity is reduced, and the applicable range is wider. Description of the Drawings

[0041] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for the description of the present invention will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 is a flowchart of the steps of a state of charge estimation method for a battery provided by an embodiment of the present invention;

[0043] Figure 2 is a diagram of a Hammerstein battery model based on an extreme learning machine provided by an embodiment of the present invention;

[0044] Figure 3It is a graph showing the change of the SOC estimated value with k provided by an embodiment of the present invention.

[0045] Figure 4 It is a graph showing the change of the estimation error with k provided by an embodiment of the present invention.

[0046] Figure 5 It is a graph showing the change of the SOC estimated value with k under the LA92 condition provided by an embodiment of the present invention.

[0047] Figure 6 It is a graph showing the change of the SOC estimated value with k under the UDDS condition provided by an embodiment of the present invention.

[0048] Figure 7 It is about different σ provided by an embodiment of the present invention 2 and under the L condition, the performance indexes of the SOC estimation curve based on the Hammerstein battery model under the HPC, LA92 and UDDS working conditions. Detailed implementation manners

[0049] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0050] The state of charge (SOC) of a battery is a key indicator for evaluating whether the battery is in an overcharged or over-discharged state and ensuring its reliable application. However, the SOC reflects the internal state of the battery rather than a specific physical quantity, so it cannot be directly measured.

[0051] Traditional experimental testing methods rely on estimation equations between the SOC and battery measurement variables to calculate the SOC. These methods mainly include the ampere-hour integration (AHI) method, the open-circuit voltage (OCV) method, and the electrochemical impedance spectroscopy (EIS) method. The ampere-hour integration (AHI) method estimates the SOC in real time by integrating the charging and discharging current of the battery, but its accuracy is affected by the initial SOC and real-time health status of the battery. The open-circuit voltage (OCV) method estimates the SOC by measuring the open-circuit voltage of the battery and locating it on the OCV-SOC characteristic curve, which is usually obtained through laboratory off-line sampling. Although the electrochemical impedance spectroscopy (EIS) method is theoretically feasible, it has limitations in practical applications.

[0052] Model-based methods are mainly divided into two categories: circuit-based models (ECM) and mechanism-based models (EM). The former simulates the charge and discharge characteristics of lithium-ion batteries through different combinations of voltage sources, resistors, and capacitors. However, such models are usually offline models and lack adaptability, so it is difficult to ensure sufficient accuracy in dynamic actual working scenarios. The latter is based on electrochemical or physical models and uses partial differential equations to describe the reaction characteristics of the battery. Such models can accurately estimate physical quantities, but due to the extremely complex internal chemical reaction mechanism of the battery, the computational cost is extremely high, resulting in limitations in practical applications.

[0053] On the other hand, data-driven methods extract features from data such as the voltage, current, and temperature of the battery, and use artificial neural networks, support vector machines, fuzzy logic, and various composite methods to fit the nonlinear time-varying characteristics of the battery, thereby constructing a battery model. The advantage of this method is that it does not need to consider the complex electrochemical reaction process inside the battery, avoiding the complexity of solving a large number of partial differential equations in the electrochemical model. However, data-driven battery models usually require a large amount of measured data for training, and this process is time-consuming.

[0054] Referring to Figure 1 , a flowchart of the steps of a method for estimating the state of charge of a battery according to an embodiment of the present invention is shown, and specifically may include the following steps:

[0055] Step 101, construct a Hammerstein battery model based on an extreme learning machine. The Hammerstein battery model includes an extreme learning machine nonlinear block and a dynamic linear block.

[0056] Single-hidden layer feedforward neural networks have good learning capabilities. Compared with traditional training methods, the extreme learning machine (ELM) algorithm has the characteristics of fast learning speed and good generalization performance. Moreover, the block-structured nonlinear model structure is flexible and can be applied to different nonlinear control technologies. Compared with the gradient algorithm based on neural networks, the algorithm steps of the extreme learning machine (ELM) are simpler because only the output weights need to be identified.

[0057] In some embodiments of the present invention, the Hammerstein model based on the extreme learning machine (ELM) uses an extreme learning machine nonlinear block followed by a linear block to perform battery modeling. Specifically, the Hammerstein battery model constructed based on the extreme learning machine is:

[0058] A(q)SOC(k) = B(q)s(k) + D(q)v(k) (1)

[0060]

[0061] where \(U(k)=[u 1 (k),u 2 (k)] T =[V(k),I(k)] T \(\in R 2 is composed of the battery terminal voltage and terminal current. The SOC of the battery is the output SOC(k). D(q)v(k) is the output of an MA model driven by white noise v(k). The linear module is a controlled autoregressive moving average (CARMA) model. s(k) is the nonlinear block of the extreme learning machine neural network. W o i is the output layer weight of the extreme learning machine. W h ji (j = 1, 2, i = 1, 2, ..., l) are the hidden layer weights of the extreme learning machine. b i is the threshold of the i-th hidden node. g(*) is the activation function. A(q), B(q) and D(q) are polynomials of known order with the backward shift operator q -1 and are defined as:

[0062] A(q): = 1 + a 1 q -1 +... + a n q -n ,

[0063] B(q): = b 0 + b 1 q -1 +... + b n q -n ,

[0064] D(q): = 1 + d 1 q -1 +... + d n q -n .

[0065] Assume that when k ≤ 0, SOC(k) = 0, U(k) = 0, v(k) = 0. The Hammerstein battery model based on the extreme learning machine (ELM) is as Figure 2 shown.

[0066] Step 102: Separate the parameters of the nonlinear block in the Hammerstein battery model from the parameters of the dynamic linear block, identify the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, and identify the output weights of the extreme learning machine nonlinear block using the least squares method.

[0067] It can be understood that the Extreme Learning Machine (ELM) algorithm randomly initializes the input weights and the hidden layer biases, and only the weights of the output layer need to be estimated. It is not difficult to see from formulas (1) and (2) that there are coupling parameters between the output weights of the non-linear block of the extreme learning machine and the parameters of the linear block. Therefore, in the embodiments of the present invention, a complex non-linear optimization problem also needs to be solved.

[0068] Specifically, according to formula (2), the non-linear block s(k) of the extreme learning machine can be written in the following matrix form:

[0069] H = g[W h U(k)+b], (3)

[0070] s(k) = H T W o , (4)

[0071] where the activation function g(*) and the input vector U(k) can be written as:

[0072]

[0073] The number of input neurons N = 2, and the input weights W h and the hidden layer biases b are randomly generated according to the number of input neurons N and the number of hidden layer neurons L, and

[0074]

[0075] The output weights W o ∈R L×M , M = 1 can be written as

[0076]

[0077] According to formula (1), to solve the identification difficulty caused by the cross product terms of b η (η = 0, 1,..., n) and W o i in the embodiments of the present invention, s(k) can be used as the key term. At the same time, for the convenience of identification, let b 0 = 1, then the linear module of the Hammerstein battery model can be written in the following form:

[0078]

[0079] Substituting the extreme learning machine formula (4) into s(k) in formula (5) can obtain:

[0080]

[0081] where the parameter vector θ and the information vector are defined as:

[0082] θ := [a 1 , a 2 ,..., a n , b 1 , b 2 ,..., b n , d 1 , d 2 ,..., d n T ∈ R 3n

[0083]

[0084] The parameter vector W o and θ can be fitted to the SOC data of the battery through formula (6). According to the hierarchical identification principle, formula (6) can be decomposed into the following two sub-identification systems:

[0085]

[0086] where, F 1 (*) and F 2 (*) are both functions, and the coupling parameters W o and θ are in two different sub-identification models. Thus, the parameters W o of the ELM nonlinear block and the parameter θ = [a 1 ,... a n , b 1 ,... b n , d 1 ,... d n T are decoupled.

[0087] The embodiment of the present invention can also define the following two criterion functions for the weights W o of the ELM nonlinear module and the linear parameter θ:

[0088]

[0089] Let and be s, v, the estimates of W o and θ at time k. For any fixed W o , applying the negative gradient search to solve the optimization problem of formula (8) constitutes the stochastic gradient (SG) algorithm to calculate θ:

[0090]

[0091] where, ​​is the innovation, is the information vector containing the historical SOC, the historical non - linear block output, and the noise, and γ(k) is the iteration step size.

[0092] For any fixed θ, the least - squares principle is applied to solve the optimization problem of formula (7) to obtain the least - squares (LS) method for calculating W o , that is:

[0093]

[0094] where, H + is the Moore - Penrose generalized inverse of the hidden - layer output matrix H, and the parameter vectors and are estimated interactively. depends on while depends on Therefore, by combining the SG method and the LS method, the SG - LS algorithm can be obtained. It can be understood that the stochastic gradient algorithm has a small computational amount and a slow convergence speed. A forgetting factor λ can be introduced to improve the convergence speed, and a convergence factor ∈ can be introduced to improve the parameter - estimation accuracy.

[0095] Furthermore, in the embodiments of the present invention, in order to further improve the parameter - estimation accuracy, the maximum - likelihood principle can also be introduced into the SG - LS algorithm to form the MLSG - LS algorithm. Let be the maximum - likelihood estimate of θ. According to the maximum - likelihood principle, the maximum - likelihood estimate can be obtained by minimizing the following criterion function:

[0096]

[0097] where,

[0098]

[0099] Define the filtering information vector:

[0100]

[0101] According to formula (12), the following expression can be obtained:

[0102]

[0103] Therefore can be written as:

[0104]

[0105] Write J 3 (θ,k) in a recursive form:

[0106]

[0107] Apply negative gradient search to minimize J 3 (θ,k), the maximum likelihood gradient algorithm can be obtained as follows:

[0108]

[0109] The gradient of can be written as:

[0110]

[0111] Therefore, the MLSG algorithm can be obtained as:

[0112]

[0113] Step 103, input the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the state of charge of the battery.

[0114] It can be understood that in practical applications, problems such as instantaneous disappearance or abnormal jump (such as power failure, signal interference) of the terminal voltage or terminal current collected by the sensor, and different dimensions of voltage and current may be faced. Direct input into the model may lead to weight deviation and other problems.

[0115] Therefore, in the embodiments of the present invention, before inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model, the terminal voltage v(k) and terminal current I(k) of the battery can also be normalized to improve the accuracy and robustness of the estimation result.

[0116] It should be noted that the embodiments of the present invention can also perform segmented dynamic adjustment processing on the normalized terminal voltage and terminal current, that is, divide the data segments according to the real-time working condition parameters of the battery, and independently optimize the normalization parameters for each data segment. Among them, the working condition parameters include at least one of the temperature change rate, charge and discharge switching state, and load fluctuation intensity. It can be understood that the battery exhibits significant differences in dynamic characteristics under different working conditions (such as charge and discharge switching, temperature change, load fluctuation). Therefore, the embodiments of the present invention can perform segmented preprocessing on the terminal voltage and terminal current. The segmented processing divides the continuous data into multiple segments according to specific rules and optimizes the processing strategies respectively to improve the adaptability of the model to complex scenarios. The segmented processing significantly improves the adaptability and robustness of the ELM-Hammerstein model by dynamically adapting to the battery working conditions. The core lies in flexible segmentation, parameter optimization and smooth transition, which not only retains the global consistency but also takes into account the local characteristics, providing an extensible solution for battery management in complex scenarios.

[0117] Further, the preprocessed voltage and current are combined into an input vector U(k), which is then input into the battery model processed in step 102 to estimate the final SOC, i.e., the state of charge of the battery.

[0118] To verify the feasibility of the above method, the present invention also provides the following embodiments:

[0119] Embodiment A: The first 80% of the hybrid pulse power characteristic (HPPC) test data of the lithium cobalt oxide battery at 25 °C is used for identification. The number of neurons is 30, and the mean absolute error (MSE) δ and root mean square error (RMSE) δ are used to evaluate the performance of the Hammerstein battery model based on the extreme learning machine (ELM). The test metrics of the SG-LS and MLSG-LS algorithms are shown in Table 1. The curves of the state of charge (SOC) estimation and error of the battery varying with k are as Figure 3 (Curve of the SOC estimated value varying with k) and Figure 4 (Curve of the estimation error varying with k) shown. It is not difficult to see that the parameter estimation accuracy of the MLSG-LS algorithm is higher than that of the SG-LS algorithm. Among them, the mean absolute error (MSE) δ, root mean square error (RMSE) δ, and goodness of fit (R2) are used to evaluate the SOC estimation performance of the Hammerstein battery model based on ELM, and the formulas are as follows:

[0120]

[0121] Among them, SOC’ is the mean value of SOC(k).

[0122] Algorithms RMSE(%) MAE(%) SG-LS 0.056687 0.0013586 MLSG-LS 0.0555 0.0011981

[0123] Table 1: Performance comparison of different algorithms

[0124] Embodiment B: Using the discharge data at 10 °C of Los Angeles 92 (LA92) and the data under the Urban Dynamometer Driving Schedule (UDDS) conditions, the SOC estimation is performed by the MLSG-LS algorithm. The curves of the SOC estimation varying with k are as Figure 5 (Curve of the SOC estimated value varying with k under LA92 conditions) and Figure 6 (Curve of the SOC estimated value varying with k under UDDS conditions) shown. The performance comparison of the proposed model with other classical neural networks under LA92 conditions is shown in Table 2.

[0125] Models RMSE(%) MAE(%) BP 2.04 1.39 LSTM 2.10 1.37 NN-Hammerstein 0.15089 0.12197 ELM-Hammerstein 0.052469 0.0008947

[0126] Table 2: Performance comparison of different models

[0127] At different σ 2Under the WLTC and UDDS conditions, the performance indicators of the SOC estimation curve based on the Hammerstein battery model under the HPC, LA92, and UDDS working conditions are as Figure 7 shown.

[0128] In an embodiment of the present invention, a Hammerstein battery model is constructed based on an extreme learning machine. The Hammerstein battery model includes a series connection of an extreme learning machine non-linear block and a dynamic linear block; the parameters of the non-linear block in the Hammerstein battery model are separated from the parameters of the dynamic linear block, and the parameters of the dynamic linear block are identified using the maximum likelihood stochastic gradient algorithm, and the output weights of the extreme learning machine non-linear block are identified using the least squares method; the terminal voltage and terminal current of the battery are input into the Hammerstein battery model to estimate the state of charge of the battery, improving the estimation accuracy of the state of charge of the battery, reducing the computational complexity at the same time, and having a wider application range.

[0129] An embodiment of the present invention also provides a state of charge estimation device for a battery, which may specifically include the following modules:

[0130] A construction module for constructing a Hammerstein battery model based on an extreme learning machine. The Hammerstein battery model includes a series connection of an extreme learning machine non-linear block and a dynamic linear block;

[0131] A parameter processing module for separating the parameters of the non-linear block in the Hammerstein battery model from the parameters of the dynamic linear block, and identifying the parameters of the dynamic linear block using the maximum likelihood stochastic gradient algorithm, and identifying the output weights of the extreme learning machine non-linear block using the least squares method;

[0132] An estimation module for inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the state of charge of the battery.

[0133] Optionally, the Hammerstein model constructed based on the extreme learning machine is:

[0134] A(q)SOC(k) = B(q)s(k) + D(q)v(k)

[0135]

[0136] Where U(k) = [u 1 (k), u 2 (k)] T = [V(k), I(k)] T ∈ R 2It is composed of the battery terminal voltage and terminal current. The SOC of the battery is the output SOC(k), D(q)v(k) is the output of an MA model driven by white noise v(k), the linear module is a controlled autoregressive moving average (CARMA) model, s(k) is the extreme learning machine neural network nonlinear block, and W o i is the weight of the output layer of the extreme learning machine, and W h ji (j = 1, 2, i = 1, 2,..., l) are the weights of the hidden layer of the extreme learning machine, and b i is the threshold of the i-th hidden node, g(*) is the activation function, and A(q), B(q), and D(q) are polynomials with known orders of the backward shift operator q -1 and are defined as:

[0137] A(q):=1 + a 1 q -1 +... + a n q -n ,

[0138] B(q):=b 0 + b 1 q -1 +... + b n q -n ,

[0139] D(q):=1 + d 1 q -1 +... + d n q -n 。

[0140] Optionally, the matrix form of the extreme learning machine nonlinear module s(k) is as follows:

[0141] H=g[W h U(k)+b],

[0142] s(k)=H T W o ,

[0143] where the activation function g(*) and the input vector U(k) are as follows:

[0144]

[0145] Optionally, the linear module of the Hammerstein battery model is as follows:

[0146]

[0147] Optionally, the parameter processing module includes:

[0148] A decoupling sub-module, which is used to decouple the parameter W of the extreme learning machine non-linear block through key item separation technology o from the parameter θ = [a 1 ,...a n ,b 1 ,...b n ,d 1 ,...d n of the dynamic linear block T .

[0149] A dynamic linear parameter identification sub-module, which is used to identify the dynamic linear block parameters by using the maximum likelihood stochastic gradient algorithm. Specifically, for the dynamic linear block parameters θ, the following maximum likelihood stochastic gradient algorithm is used for iterative update:

[0150]

[0151] where is an information vector containing historical SOC, historical non-linear block output and noise, and γ(k) is the iteration step size

[0152] A non-linear block output weight identification sub-module, which is used to identify the extreme learning machine non-linear block output weight by using the least squares method. Specifically, for the extreme learning machine non-linear block output weight W o , the following least squares method is used for calculation:

[0153]

[0154] where H is the hidden layer output matrix, and H + is its Moore-Penrose generalized inverse

[0155] Optionally, before inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model, it further includes:

[0156] Normalizing the terminal voltage and terminal current

[0157] Optionally, the method further includes:

[0158] Dividing the data segments according to the real-time working condition parameters of the battery, and independently optimizing the normalization parameters for each data segment, where the implementation working condition parameters include at least one of the temperature change rate, charge and discharge switching state, and load fluctuation intensity

[0159] It should be noted that, for method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the described action sequence, because according to the embodiments of the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential for the embodiments of the present invention.

[0160] For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiments.

[0161] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. And the collection, use, and processing of the relevant data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entrances are provided for the user to choose to authorize or refuse.

[0162] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other.

[0163] The above provides a detailed introduction to a method and apparatus for estimating the state of charge of a battery. Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for estimating a battery state of charge, characterized in that: The method comprises: Building a Hammerstein battery model based on an extreme learning machine, wherein the Hammerstein battery model includes a series connection of an extreme learning machine nonlinear block and a dynamic linear block; Separating the parameters of the nonlinear block from the parameters of the dynamic linear block in the Hammerstein battery model, identifying the parameters of the dynamic linear block using a maximum likelihood stochastic gradient algorithm, and identifying the output weights of the extreme learning machine nonlinear block using a least squares method; The terminal voltage and terminal current of the battery are input into the Hammerstein battery model to estimate the battery state of charge.

2. The method according to claim 1, characterized in that The Hammerstein model constructed based on the extreme learning machine is: A(q)SOC(k)=B(q)s(k)+D(q)v(k) Where U(k)=[u1(k),u2(k)] T =[V(k),I(k)] T ∈R 2 It is composed of the battery terminal voltage and terminal current. The battery SOC is output SOC(k). D(q)v(k) is the output of the MA model driven by white noise v(k). The linear module is a controlled autoregressive moving average (CARMA) model. s(k) is the extreme learning machine neural network nonlinear block. W o i is the output layer weight of the extreme learning machine, W h ji (j=1,2,i=1,2,...,l) is the hidden layer weight of the extreme learning machine, b i is the threshold of the i-th hidden node, g(*) is the activation function, A(q), B(q) and D(q) are the back-shift operators for q -1 A polynomial of known order is defined as: A(q):=1+a1q -1 +...+a n q -n , B(q):=b0+b1q -1 +...+b n q -n , D(q):=1+d1q -1 +...+d n q -n 。 3. The method according to claim 2, characterized in that The matrix form of the extreme learning machine nonlinear module s(k) is as follows: H=g[W h U(k)+b], s(k)=H T W o , Among them, the activation function g(*) and the input vector U(k) are as follows:

4. The method according to claim 3, characterized in that: The linear module of the Hammerstein battery model is as follows:

5. The method according to claim 4, characterized in that The separating of the parameters of the nonlinear block from the parameters of the dynamic linear block in the Hammerstein battery model comprises: The parameters W of the nonlinear block of the extreme learning machine are separated by the key term separation technique. o and the parameters of the dynamic linear block θ=[a1,...a n ,b1,...b n ,d1,...d n ] T Among them, b i Decoupling.

6. The method according to claim 5, characterized in that The method of identifying the dynamic linear block parameters by using the maximum likelihood stochastic gradient algorithm includes: The dynamic linear block parameter θ is iteratively updated using the following maximum likelihood stochastic gradient algorithm: in, is the information vector containing historical SOC, historical nonlinear block output and noise, and γ(k) is the iteration step size.

7. The method according to claim 6, characterized in that The method of using the least squares method to identify the output weight of the nonlinear block of the extreme learning machine includes: Output weight W for the extreme learning machine nonlinear block o , calculated using the following least squares method: Among them, H is the hidden layer output matrix, H + is its Moore-Penrose generalized inverse.

8. The method according to claim 1, characterized in that Before inputting the terminal voltage and terminal current of the battery into the Hammerstein battery model, it also includes: The terminal voltage and the terminal current are normalized.

9. The method according to claim 8, characterized in that The method further comprises: The data segments are divided according to the real-time operating condition parameters of the battery, and normalization parameters are independently optimized for each segment of data, wherein the operating condition parameters include at least one of a temperature change rate, a charge and discharge switching state, and a load fluctuation intensity.

10. A battery state of charge estimation device, characterized in that: The device comprises: A building module for building a Hammerstein battery model based on an extreme learning machine, wherein the Hammerstein battery model includes a series connection of an extreme learning machine nonlinear block and a dynamic linear block; A parameter processing module, used to separate the parameters of the nonlinear block from the parameters of the dynamic linear block in the Hammerstein battery model, identify the parameters of the dynamic linear block using a maximum likelihood stochastic gradient algorithm, and identify the output weights of the nonlinear block of the extreme learning machine using a least squares method; The estimation module is used to input the terminal voltage and terminal current of the battery into the Hammerstein battery model to estimate the battery state of charge.

Images