Online prediction method and system for state of charge of lithium-ion batteries

Abstract

The present invention belongs to the field of lithium-ion battery technology optimization, and in particular, relates to an online prediction method and system for the state of charge of a lithium-ion battery, including: establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and using the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data; establishing a lithium-ion battery state of charge offline prediction model based on the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery; on the basis of the lithium-ion battery state of charge offline prediction model, adding nodes for new input data, establishing a lithium-ion battery state of charge online prediction model, and predicting the state of charge of the lithium-ion battery through the lithium-ion battery state of charge online prediction model. The present invention combines the spatiotemporal modeling strategy and width learning to predict the state of charge of the lithium battery, and improves the prediction efficiency of the model while ensuring the calculation speed and output accuracy of the model.

Description

Technical Field

[0001] The present invention belongs to the field of lithium-ion battery technology optimization, and in particular relates to an online prediction method and system for the state of charge of a lithium-ion battery. Background Art

[0002] The battery state of charge (SOC) is a very important parameter indicator in the operation of electric vehicles. It is an important basis for judging the remaining battery power, preventing the battery from overcharging and over-discharging, and improving battery performance. It is also one of the key technologies that the battery management system needs to solve. Similar to the fuel gauge of a traditional fuel vehicle, the battery SOC reflects the remaining battery power. However, unlike the way fuel vehicles detect the remaining fuel, the battery SOC is an internal parameter of the battery and cannot be directly measured by various sensors. Since the battery SOC is closely related to other external parameters that can be directly measured, the SOC can be indirectly estimated through other measured parameters, such as battery terminal voltage, charge and discharge current, battery temperature, etc.

[0003] However, in some cases, it is difficult to install one or more temperature sensors on the surface of lithium-ion batteries to obtain temperature data. At the same time, the temperature of lithium batteries has a great influence on the SOC of the battery. Therefore, establishing an accurate temperature field prediction model is conducive to the prediction of lithium battery SOC. The temperature field of lithium-ion batteries belongs to a type of time-varying distributed parameter system. This type of system has time-varying and strong nonlinear characteristics. The system input and output, system boundaries and system parameters will change with time and space. On the other hand, the internal state of lithium batteries under different load conditions is also different. When the accuracy is insufficient, it is necessary to be able to update the structure of the SOC prediction model in time. Therefore, it is a huge challenge in related industries to establish an accurate temperature prediction model for lithium batteries and further establish a fast and effective SOC prediction model.

[0004] In the prior art, the SOC estimation methods for lithium-ion batteries mainly include the ampere-hour integration method, the open circuit voltage method, the internal resistance method, the physical model method, the data-driven method and the fusion algorithm.

[0005] Among them, experimental methods such as the ampere-hour integration method, the open circuit voltage method, the internal resistance method, etc., need to rely on specific instruments or operating steps, and the entire operation process is time-consuming. The SOC estimation result is greatly affected by the measurement accuracy, so it is difficult to apply in actual situations;

[0006] The estimation method based on physical models mainly links the external parameters of lithium-ion batteries with internal state variables, establishes mathematical relationships, and estimates the internal state variables of lithium-ion batteries in combination with algorithms such as statistical filtering. Xing Y et al. established a first-order RC equivalent circuit model and obtained multiple sets of OCV-SOC curves through discharge experiments at different temperatures. The state equation and measurement equation were established, and the UKF algorithm was combined to accurately estimate the SOC at different temperatures. However, the model is relatively complex and the model parameters need to be adjusted continuously with the temperature, which requires a large amount of calculation.

[0007] The data-driven method regards the lithium-ion battery as a black box, no longer focusing on its internal working principle and reaction mechanism, and achieves the estimation of SOC by establishing a data-driven model to describe the nonlinear relationship between the external parameters of the battery and SOC. Currently, the commonly used data-driven methods include back propagation (BP) neural network, radial basis function (RBF) neural network, extreme learning machine (ELM), support vector machine (SVM), etc.

[0008] Fusion algorithms mainly refer to algorithms that integrate data-driven and statistical filtering algorithms, which have higher estimation accuracy and stability. Tian Y et al. proposed a long short-term memory (LSTM) recursive neural network to simulate the behavior of lithium-ion batteries at different temperatures, and estimated the battery SOC from voltage, current and temperature variables, and introduced an unscented Kalman filter to filter out noise and further reduce the estimation error. The disadvantage of the two types of technologies, such as data-driven methods and fusion algorithms, is that when the prediction accuracy is insufficient, adjusting the network structure requires retraining.

[0009] In the above technology, the temperature data for estimating the SOC of lithium-ion batteries comes from the temperature sensor installed on the surface of lithium-ion batteries. However, during the use of lithium-ion batteries, current, voltage and other data can be directly measured through their external circuits. However, in some use cases, due to space and cost constraints, it is impossible to install one or more temperature sensors on lithium-ion batteries to obtain temperature data, which in turn limits the further improvement of the battery SOC estimation accuracy. In addition, for the time-varying behavior of lithium-ion battery systems, the model needs to adapt to changes in future load conditions in a timely manner. When the network accuracy is insufficient, adjusting the network structure requires retraining, which takes a certain amount of time and cannot provide timely and accurate prediction performance. Summary of the invention

[0010] The present invention aims to solve the problem in the prior art that it is impossible to install one or more temperature sensors on lithium-ion batteries to obtain temperature data, and the constructed charge state prediction model cannot adapt to changes in future load conditions.

[0011] To solve the above problems, in a first aspect, an embodiment of the present invention provides an online prediction method for the state of charge of a lithium-ion battery, the online prediction method comprising the following steps:

[0012] Establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and using the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data;

[0013] Establishing an offline prediction model for the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery;

[0014] On the basis of the offline prediction model of the state of charge of the lithium-ion battery, nodes are added for new input data to establish an online prediction model of the state of charge of the lithium-ion battery, and the state of charge of the lithium-ion battery is predicted by the online prediction model of the state of charge of the lithium-ion battery to obtain a prediction result.

[0015] Furthermore, the step of establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy comprises the following sub-steps:

[0016] Performing dimensionality reduction processing on the spatiotemporal data using a KPCA algorithm to obtain a spatial basis function that satisfies a preset number of spatial basis functions, and obtaining a corresponding low-order time coefficient;

[0017] Using the ELM model to learn the low-order time coefficients, to obtain a low-order time series model with time dynamic characteristics;

[0018] The subsequent time coefficients are predicted using a low-order time series model, and then reconstructed in conjunction with the spatial basis function to obtain the lithium-ion battery temperature prediction model, and the temperature spatiotemporal prediction data is predicted by the lithium-ion battery temperature prediction model.

[0019] Furthermore, the spatiotemporal data is defined as y(x, t), and the spatiotemporal data satisfies the following relationship:

[0020]

[0021] In the above formula, represents the spatial basis function, a i (t) represents the low-order time coefficient, and n is the number of components that can represent the entire spatiotemporal data calculated by the KPCA algorithm;

[0022] The low-order timing model is modeled according to the low-order time coefficients through the ELM model, wherein the low-order time coefficients used by the ELM model satisfy:

[0023] a i (t) = f(a i (t-1),...,a i (td a ),u(t-1),...,u(td u ))+ε(t);

[0024] Among them, d a and d u Represent the maximum delay of input signal and output signal respectively;

[0025] The lithium-ion battery temperature prediction model satisfies:

[0026]

[0027] in, The temperature spatiotemporal prediction data obtained by prediction.

[0028] Furthermore, the step of establishing an offline prediction model for the state of charge of a lithium-ion battery according to the spatiotemporal prediction data and the voltage and current data of the lithium-ion battery comprises the following sub-steps:

[0029] Inputting the current data, the voltage data, and the temperature spatiotemporal prediction data as input variables into a preset width learning network for training;

[0030] The preset width learning network that has completed training is output as the offline prediction model for the state of charge of the lithium-ion battery.

[0031] Furthermore, the preset width learning network satisfies:

[0032]

[0033] Wherein, Y is the prediction result output by the preset width learning network, Z n represents the mapping feature node when the current data, the voltage data, and the temperature spatiotemporal prediction data are input as input variables into the preset width learning network, Z n represents the set of all the mapping feature nodes, H m represents the enhanced node corresponding to the mapping feature node, H m represents the set of all the enhanced nodes, and are randomly generated weights and biases, ξ represents the activation function, Wm is the connection weight of the width structure.

[0034] Furthermore, based on the offline prediction model for the state of charge of the lithium-ion battery, a node is added for new input data to establish an online prediction model for the state of charge of the lithium-ion battery, and the state of charge of the lithium-ion battery is predicted by the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result, including the following sub-steps:

[0035] Acquire the new input data, add nodes to the offline prediction model of the state of charge of the lithium-ion battery, and calculate the increment according to the input data;

[0036] Calculating the updated weight of the node according to the incremental calculation of the input data;

[0037] Applying the updated weight to the lithium-ion battery state of charge offline prediction model, and outputting the lithium-ion battery state of charge online prediction model;

[0038] The state of charge of the lithium-ion battery is predicted by using the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result of the state of charge.

[0039] Furthermore, the input data is defined as X a , and Representing n groups of the mapping feature nodes and m groups of the enhancement nodes in the offline prediction model for the state of charge of the lithium-ion battery, the increment satisfies:

[0040]

[0041] in, By X a Updated incremental feature set, W e1 ,W en and β e1 ,β e1 Randomly generated during the initial operation of the network;

[0042] The update weights satisfy:

[0043]

[0044] where Y a Is additional X a The corresponding label of .

[0045] In a second aspect, an embodiment of the present invention further provides an online prediction system for the state of charge of a lithium-ion battery, comprising:

[0046] A temperature prediction module is used to establish a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and use the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data;

[0047] An offline prediction module is used to establish an offline prediction model of the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery;

[0048] The online prediction module is used to add nodes for new input data based on the offline prediction model of the lithium-ion battery state of charge to establish an online prediction model of the lithium-ion battery state of charge, and predict the state of charge of the lithium-ion battery through the online prediction model of the lithium-ion battery state of charge to obtain a prediction result.

[0049] The beneficial effect achieved by the present invention is that the method for predicting the state of charge of a lithium battery by combining a spatiotemporal modeling strategy and width learning is adopted. The temperature of the lithium battery can be predicted directly according to the measured values ​​of an external circuit, such as current and voltage. The input data set is further used to predict the state of charge of the lithium battery d1 through width learning. This can be applied to situations where it is difficult to install one or more temperature sensors on the lithium battery to obtain temperature data. At the same time, the online prediction model of the state of charge of a lithium battery established based on width learning has good scalability brought by the width learning network. When the prediction accuracy is insufficient, the network structure can be optimized by increasing the number of enhanced nodes. This improves the prediction efficiency of the model while ensuring the calculation speed and output accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the steps of an online prediction method for the state of charge of a lithium-ion battery provided by an embodiment of the present invention;

[0051] Figure 2 Schematic diagram of a modeling framework of a lithium-ion battery temperature prediction model provided by an embodiment of the present invention;

[0052] Figure 3 is a schematic diagram of the ELM model structure provided by an embodiment of the present invention;

[0053] Figure 4 It is a schematic diagram of the incremental learning algorithm framework provided by an embodiment of the present invention;

[0054] Figure 5 It is a schematic diagram of an online prediction model for the state of charge of a lithium-ion battery provided by an embodiment of the present invention;

[0055] Figure 6 is a schematic diagram of the distribution of lithium battery surface temperature sensors provided by an embodiment of the present invention;

[0056] Figure 7 is the sensor temperature test result provided by the embodiment of the present invention;

[0057] Figure 8 is a schematic diagram of a state of charge test result provided by an embodiment of the present invention;

[0058] Fig. 9 The figure is a schematic diagram of the structure of an online prediction system for the state of charge of a lithium-ion battery provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] Please refer to Figure 1 , Figure 1 1 is a flowchart of the steps of an online prediction method for the state of charge of a lithium-ion battery provided by an embodiment of the present invention, wherein the online prediction method comprises the following steps:

[0061] S1. A lithium-ion battery temperature prediction model is established based on a spatiotemporal modeling strategy, and the spatiotemporal data of the lithium-ion battery temperature is predicted using the lithium-ion battery temperature prediction model to obtain temperature spatiotemporal prediction data.

[0062] Furthermore, the step of establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy comprises the following sub-steps:

[0063] Performing dimensionality reduction processing on the spatiotemporal data using a KPCA algorithm to obtain a spatial basis function that satisfies a preset number of spatial basis functions, and obtaining a corresponding low-order time coefficient;

[0064] Using the ELM model to learn the low-order time coefficients, to obtain a low-order time series model with time dynamic characteristics;

[0065] The subsequent time coefficients are predicted using a low-order time series model, and then reconstructed in conjunction with the spatial basis function to obtain the lithium-ion battery temperature prediction model, and the temperature spatiotemporal prediction data is predicted by the lithium-ion battery temperature prediction model.

[0066] Specifically, the modeling framework of the lithium-ion battery temperature prediction model is as follows: Figure 2 As shown. The temperature distribution of the lithium battery temperature field belongs to a typical nonlinear distributed parameter system. Mathematically, the distributed parameter system can be described using partial differential equations PDE. Since such systems have the characteristics of spatial distribution, they are essentially infinite-dimensional systems. For ease of understanding, the PDE description of the distributed parameter system can be expressed as follows:

[0067]

[0068] Among them, α, β and ω are constants; u(t) is the control value; b(x) is the spatial distribution matrix of u(t).

[0069] The boundary conditions are:

[0070] y(0,t)=0,y(π,t)=0 (2);

[0071] The initial conditions are:

[0072] y(x,0)=y0(x) (3);

[0073] According to the Fourier transform, any nonlinear continuous function can be expanded using the Fourier series. Therefore, the space-time variable y(x, t) can be expanded along a set of spatial basis functions To perform projection expansion:

[0074]

[0075] in represents the spatial basis function, a i (t) represents the corresponding low-order time coefficient.

[0076] In practical applications, formula (4) can generally be expanded into an approximate finite term form, namely:

[0077]

[0078] Where n represents the number of spatial basis functions that can contain most of the spatial information.

[0079] In order to find the most typical or characteristic spatial basis function of the system Solve the optimization problem of minimizing the objective function:

[0080]

[0081] In the above formula, y n (x, t) is the process output obtained from the experiment distributed in time and space; is the L2 norm of f(x); is its ensemble mean; (f(x)·g(x))=∫ Ω f(x)g(x)dx represents the inner product of f(x) and g(x); L 2 (Ω) represents the L2 norm space; spatial basis function Under the given constraints is unique in . Applying the Lagrange multiplier method, we can get:

[0082]

[0083] The necessary condition for its extreme value can be expressed as:

[0084]

[0085] Where R(x,ζ)=<y(x,t)y(ζ,t)> is the correlation function of two points in space; is the spatial basis function of the variable ζ. Solving equation (8) can yield the solution to the optimization problem.

[0086] According to the snapshot method, the spatial basis function can be expressed as a linear combination of space-time variables:

[0087]

[0088] n t Represents the amount of data on a time scale.

[0089] Substituting (9) into (8) yields the following eigenvalue problem:

[0090]

[0091] The correlation function between two points in time is defined as follows:

[0092]

[0093] C tk is the correlation function between two points in time.

[0094] Therefore, the eigenvalue problem in equation (10) is transformed from the n×n dimension into the following n t ×n t dimension:

[0095] Cγ i =λ i γ i (12);

[0096] in is the ith eigenvector.

[0097] By solving (12), we can obtain the eigenvector and the corresponding eigenvalues Then, the orthogonal space basis function is constructed by (9): It is orthogonal, and the spatial basis function can be obtained by standardizing it Therefore, the process of learning the spatial basis functions becomes learning γ ik process.

[0098] In the embodiment of the present invention, the KPCA algorithm is used to learn the spatial basis functions.

[0099] KPCA through nonlinear mapping It realizes the transformation from input space to feature space, and then performs linear PCA on the mapped data, thus having strong linear processing capabilities;

[0100] For M samples x in the input space k (k=1,...,M), x k ∈R N ,make Then its autocorrelation matrix is:

[0101]

[0102] For the general PCA method, that is, by solving the characteristic equation:

[0103] Cν=λν (14);

[0104] Obtain the eigenvalue with the largest contribution rate. In the embodiment of the present invention, the corresponding larger eigenvalue and the corresponding eigenvector are taken. Introduce the nonlinear feature mapping function φ to transform the sample points x1, x2, ..., xM in the input space into sample points in the feature space And assume that:

[0105]

[0106] Then the autocorrelation matrix in the feature space F is:

[0107]

[0108] The autocorrelation matrix of formula (16) Finding the eigenvalues ​​and eigenvectors gives the following formula:

[0109]

[0110] For λ≠0, all eigenvectors satisfying equation (17) are in the eigenvector In the space generated by the collection, there are two useful information that can be used:

[0111] First, the equivalent equation of formula (17):

[0112]

[0113] Second, there exists a corresponding set of coefficients that can be written as:

[0114]

[0115] Substituting equation (16) and equation (19) into equation (18), we can obtain:

[0116]

[0117] Among them, k=1,...,M.

[0118] In order to solve equation (20), observe the right side of the equation and define an N×N matrix K:

[0119]

[0120] Define two matrices:

[0121] The N×N K matrix is ​​the introduced kernel matrix, which can be calculated by substituting the data in the original space into the kernel function;

[0122] N×N vector α, where the jth element is the parameter α j .

[0123] Therefore, formula (20) can be written as follows:

[0124]

[0125] Since there is a kernel matrix K on both ends of equation (22), the problem of finding the eigenvalue of the above equation can be transformed into:

[0126]

[0127] Let represent the eigenvalue of the kernel matrix K, that is

[0128]

[0129] in Represents the correlation matrix The jth eigenvalue of . Thus, equation (23) can be transformed into the following standard form:

[0130] Kα=λα (25);

[0131] In order to obtain the eigenvector of unit length, it is normalized. According to (19) and (25), α needs to be processed as follows:

[0132]

[0133] where p is the number of non-zero eigenvalues ​​in the kernel matrix.

[0134] From formula (15), we can see that we assume that the data in the feature space is centralized, but this is not the case in reality. To obtain the kernel matrix of the normalized data, it needs to be processed as follows:

[0135]

[0136] in,

[0137] Theoretically, the more the real situation of the space-time system is reflected, the more complicated the model construction is when the entire space basis function is considered. Therefore, it is necessary to select n (n ≤ K) dominant space basis functions to approximate the main dynamics of the space-time system. Define the cumulative contribution rate:

[0138]

[0139] Where M≤(n,n t ) is the maximum number of non-zero eigenvalues. E i It represents the ratio of the sum of the n largest eigenvalues ​​to the sum of all eigenvalues.

[0140] When the obtained cumulative contribution rate meets the requirements, the first n eigencomponents can be used to represent the information of the entire eigenvector space, where the first n components in the eigenvector are also called the principal components of the eigenvector. Generally, E i >0.99999.

[0141] In the embodiment of the present invention, the ELM model is used to learn the low-order time coefficients to obtain a low-order timing model that can represent the time dynamic characteristics of the system. The specific process is as follows:

[0142] The ELM model is the extreme learning machine, which is essentially a single hidden layer feedforward neural network SLFNs. Its structure is as follows: Figure 3 shown.

[0143] Given an input-output sample set n u and n y are the number of input and output neurons, respectively, and N represents the number of measurements; there are L hidden neurons in the SLFNs output function. In the SLFNs output function, represents the output weight vector connecting the output neuron and the i-th hidden neuron, represents the input weight vector connecting the input neuron and the i-th hidden neuron, o t represents the output value of ELM at time t, η i represents the threshold of the i-th hidden layer neuron, ω i ·u t Represents ω i u t The inner product of and , G(·) is used to represent the activation function of the hidden layer neurons.

[0144] Any nonlinear continuous function f(x) can be approximated by SLFNs with zero error, that is where ||·|| represents the F-norm. Therefore, there exists ω i , η i and β i ,therefore:

[0145]

[0146] in:

[0147]

[0148]

[0149] In formula (30), the matrix H is the hidden layer output matrix of SLFNs, which represents the result of the data after the hidden layer operation, where the i-th column is the input data u1,u2,...,u N The output result of the i-th hidden layer neuron of H is the input data u j The output vector composed of all hidden layer neurons.

[0150] Before training the ELM network, the input weights ω of the model need to be i and hidden layer threshold η i Randomly generated, its input weight ω i and hidden layer threshold η i Independent of the training data, no update is required. After determining the appropriate activation function, the output weight β of the model can be obtained by the least squares method of formula (32).

[0151]

[0152] in represents the Moore–Penrose (MP) generalized inverse of the matrix H.

[0153] The ELM model learns a low-order time series model; after calculating the appropriate spatial basis function After that, the next step is to model the low-order dynamics; in general, the low-order time coefficients of nonlinear time series dynamics are expressed as follows:

[0154] a i (t) = f(a i (t-1),...,a i (td a ),u(t-1),...,u(td u ))+ε(t) (33);

[0155] Where, d a and d u Represent the maximum delay of input signal and output signal respectively;

[0156] The ELM model can be used to approximate any nonlinear system, and after selecting a suitable input signal, the output signal at the current moment can be obtained by combining the previous low-order time series output.

[0157] The prediction output of the temperature spatiotemporal prediction data is obtained by the following formula:

[0158]

[0159] in, The temperature spatiotemporal prediction data obtained by prediction.

[0160] S2. Establishing an offline prediction model for the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery.

[0161] Furthermore, the step of establishing an offline prediction model for the state of charge of a lithium-ion battery according to the spatiotemporal prediction data and the voltage and current data of the lithium-ion battery comprises the following sub-steps:

[0162] Inputting the current data, the voltage data, and the temperature spatiotemporal prediction data as input variables into a preset width learning network for training;

[0163] The preset width learning network that has completed training is output as the offline prediction model for the state of charge of the lithium-ion battery.

[0164] Specifically, BLS is constructed based on the traditional RVFLNN. However, unlike the traditional RVFLNN that directly obtains input and builds enhanced nodes, BLS first maps the input to construct a set of mapping features. In addition, the incremental learning algorithm can also dynamically update the system. Its structure is as follows Figure 4 shown.

[0165] First, the current, voltage, and temperature data are input into the width learning network as input variables. Project the data into mapping feature nodes Z i , where W ei are random weights with appropriate dimensions. G i ≡[G1,…G i ] is the concatenation of all the first i groups of mapping feature nodes. Similarly, the jth group of enhancement nodes Indicated as H j , the concatenation of all the first j groups of enhanced nodes is represented as H j ≡[H1,…,H j ]. i and j can be selected differently according to the complexity of the modeling task. In addition, φ i and φ k can be different functions of i≠k. Similarly, ξj and r It can also be a different function of j≠r. Without loss of generality, the following formula omits the i-th random mapping φ i and the jth random mapping ξ j The subscript of .

[0166] Assume that the input dataset is equipped with N samples, each sample has M dimensions, and Y belongs to The output matrix of . For n feature maps, each map generates k nodes, which can be expressed as a formal equation:

[0167]

[0168] Among them, W ei and β ei are randomly generated weights and biases; φ represents the activation function. All feature nodes are represented as Z n ≡[Z1,…Z n ], and the mth group of enhanced nodes is represented as:

[0169]

[0170] in, and are randomly generated weights and biases; ξ represents the activation function.

[0171] Therefore, the generalized model can be expressed as an equation of the form:

[0172]

[0173] Where W m =[Z n |H m ] + Y, W m is the connection weight of the width structure, and (17) is used to calculate [Z n |H m ] + The ridge regression approximation of . It can be easily calculated:

[0174]

[0175] S3. On the basis of the offline prediction model for the state of charge of the lithium-ion battery, nodes are added for new input data to establish an online prediction model for the state of charge of the lithium-ion battery, and the state of charge of the lithium-ion battery is predicted by the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result.

[0176] Furthermore, based on the offline prediction model for the state of charge of the lithium-ion battery, a node is added for new input data to establish an online prediction model for the state of charge of the lithium-ion battery, and the state of charge of the lithium-ion battery is predicted by the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result, including the following sub-steps:

[0177] Acquire the new input data, add nodes to the offline prediction model of the state of charge of the lithium-ion battery, and calculate the increment according to the input data;

[0178] Calculating the updated weight of the node according to the incremental calculation of the input data;

[0179] Applying the updated weight to the lithium-ion battery state of charge offline prediction model, and outputting the lithium-ion battery state of charge online prediction model;

[0180] The state of charge of the lithium-ion battery is predicted by using the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result of the state of charge.

[0181] Specifically, after the offline model is built, as data continues to be input, the model should be updated to reflect the additional samples, and BLS can easily update the weights without the need for a full training cycle. Its structure is as follows Figure 5 shown.

[0182] X a Represented as the new input data added to the neural network, It is represented as n groups of mapping feature nodes and m groups of enhancement nodes of the initial network. The incremental formula of mapping feature nodes and enhancement nodes is as follows:

[0183]

[0184] in By X a Updated incremental feature set; W e1 ,W en and β e1 ,β e1 It is randomly generated during the initial operation of the network. Therefore, the update matrix is ​​defined as:

[0185]

[0186] The related pseudo-inverse update algorithm can be inferred as follows:

[0187]

[0188] in

[0189]

[0190] and Therefore, the updated weights are:

[0191]

[0192] where Y a Is additional X a The corresponding label of .

[0193] By continuously inputting data into the network and using the incremental learning algorithm to quickly update the network weights, the timeliness of the data is guaranteed, thereby obtaining a fast and accurate model output SOC.

[0194] Illustratively, a group of examples implemented by the embodiment of the present invention using the above-mentioned method for online prediction of the state of charge of a lithium-ion battery are shown below.

[0195] like Figure 6 As shown, there are 32 temperature sensors evenly distributed on the surface of the lithium battery. In order to fully stimulate the battery temperature, the input current and the generated working voltage are used as the input signal of the system u(t)=[I(t),U(t)], where I(t) is the input current and U(t) is the working voltage.

[0196] When the sample interval Δt=1s, the data under the UDDS condition and the US06 condition were used, which were 22144 and 11757 groups respectively. Each group contained 32 temperature data, current data, and voltage data collected by thermal sensors.

[0197] In the initial stage, the KPCA method is used to separate the temperature data y(x, t) into spatial basis functions containing spatial information and corresponding low-order time coefficients related to time.

[0198] The eigenvector and eigenvalue are obtained by formula (12), and the corresponding eigenvectors are rearranged in the order of eigenvalue from large to small; the appropriate cumulative contribution rate E is selected as the basis for determining the number of spatial basis functions, and the cumulative contribution rates E1, ..., E1 of the first n principal components are calculated in turn according to formula (13): i , when E i When >E, the first n eigenvectors are selected as principal components; the first 32 sets of data are selected as snapshots, and the principal components obtained in the previous step are substituted into formula (9) by the snapshot method to obtain the corresponding spatial basis function Then, the discrete spatial basis functions are transformed into basis functions that are continuously distributed in space by using the cubic spline interpolation method. After obtaining the spatial basis function, the spatiotemporal data is projected onto the spatial basis function to obtain the corresponding low-order time coefficient a(t). The first 18,000 sets of data under the UDDS condition are used as training samples, and the last 4,144 sets are used as test samples.

[0199] According to the obtained low-order time coefficients, the low-order time coefficients of the first 18,000 sets of data under UDDS working conditions are used as training data to establish the ELM low-order time series model. Then, the last 4,144 sets of data are used as test set data to test the temperature prediction effect of the model. The temperatures of sensors s3, s12, and s26 are selected to test the model effect. Figure 7 shown.

[0200] In the second stage, the UDDS operating data is used as the training set of the SOC prediction model, and the US06 operating data is used as the test set of the SOC prediction model. The current, voltage, and the average temperature collected by 32 sensors are normalized, and then the training samples are input into the width learning network, and the random weight W ei and deviation β ei , set enhanced node group H m ≡[H1,…,H m ], the weight W can be calculated by formula (38) m .

[0201] The current and voltage data under the US06 working condition are used as the input signal u(t)=[I(t),U(t)] and input into the temperature prediction model of the first stage to obtain the temperature prediction data of 32 positions of the lithium battery. The temperature prediction data of the 32 temperature sensors of this predicted temperature data are averaged to obtain the average temperature data, a total of 11757 groups. This average temperature data and the current and voltage data in the US06 working condition data form the test data of the online prediction model.

[0202] After the initial model was established, 11731 sets of test data were used as newly arrived data for online learning. For the newly input data, if the accuracy is not enough, the width learning network will optimize the network structure according to its incremental learning algorithm and update the weights. According to equations (41, 42, 43), equation (40) can be solved to obtain the updated weights Thus, the SOC of the lithium battery output by the model is obtained. The test set effect is as follows Figure 8 As shown, it can be seen that the prediction result obtained by the online prediction method for the state of charge of a lithium-ion battery in the embodiment of the present invention is highly accurate and can effectively perform online prediction of the SOC of a lithium battery under different working conditions.

[0203] The beneficial effect achieved by the present invention is that the method for predicting the state of charge of a lithium battery by combining a spatiotemporal modeling strategy and width learning is adopted. The temperature of the lithium battery can be predicted directly according to the measured values ​​of an external circuit, such as current and voltage. The input data set is further used to predict the state of charge of the lithium battery d1 through width learning. This can be applied to situations where it is difficult to install one or more temperature sensors on the lithium battery to obtain temperature data. At the same time, the online prediction model of the state of charge of a lithium battery established based on width learning has good scalability brought by the width learning network. When the prediction accuracy is insufficient, the network structure can be optimized by increasing the number of enhanced nodes. This improves the prediction efficiency of the model while ensuring the calculation speed and output accuracy of the model.

[0204] The embodiment of the present invention also provides an online prediction system 200 for the state of charge of a lithium-ion battery. Fig. 9 As shown, including:

[0205] The temperature prediction module 201 is used to establish a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and use the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data;

[0206] An offline prediction module 202 is used to establish an offline prediction model of the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery;

[0207] The online prediction module 203 is used to add nodes for new input data based on the offline prediction model of the lithium-ion battery state of charge to establish an online prediction model of the lithium-ion battery state of charge, and predict the state of charge of the lithium-ion battery through the online prediction model of the lithium-ion battery state of charge to obtain a prediction result.

[0208] The lithium-ion battery state of charge online prediction system 200 can implement any one of the steps in the lithium-ion battery state of charge online prediction method in the above embodiment of the present invention when running, and can achieve the same technical effect. To avoid repetition, it will not be described here.

[0209] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through computer modeling, and the modeling can be stored in a computer-readable storage medium. When the modeling is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM). For example, in one possible implementation, the computer-readable storage medium stores a computer modeling, and when the computer modeling is executed by the processor, it implements the various processes and steps in the online prediction method of the state of charge of a lithium-ion battery provided in the embodiment of the present invention, and can achieve the same technical effect. In order to avoid repetition, it will not be repeated here.

[0210] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0211] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0212] The embodiments of the present invention are described above in conjunction with the accompanying drawings. What is disclosed is only the preferred embodiment of the present invention. However, the present invention is not limited to the above-mentioned specific implementation manner. The above-mentioned specific implementation manner is only illustrative rather than restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms and equivalent changes without departing from the scope of protection of the purpose of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for online prediction of the state of charge of a lithium-ion battery, characterized in that: The online prediction method comprises the following steps: Establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and using the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data; Establishing an offline prediction model for the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery; On the basis of the offline prediction model for the state of charge of the lithium-ion battery, adding nodes for new input data to establish an online prediction model for the state of charge of the lithium-ion battery, and performing a prediction on the state of charge of the lithium-ion battery through the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result; The step of establishing a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy includes the following sub-steps: Performing dimensionality reduction processing on the spatiotemporal data using a KPCA algorithm to obtain a spatial basis function that satisfies a preset number of spatial basis functions, and obtaining a corresponding low-order time coefficient; Using the ELM model to learn the low-order time coefficients, to obtain a low-order time series model with time dynamic characteristics; The subsequent time coefficients are predicted by using a low-order time series model, and then reconstructed in conjunction with the spatial basis function to obtain the lithium-ion battery temperature prediction model, and the temperature spatiotemporal prediction data is predicted by the lithium-ion battery temperature prediction model; The step of establishing an offline prediction model for the state of charge of a lithium-ion battery according to the spatiotemporal prediction data and the voltage and current data of the lithium-ion battery comprises the following sub-steps: Inputting the current data, the voltage data, and the temperature spatiotemporal prediction data as input variables into a preset width learning network for training; Outputting the trained preset width learning network as the offline prediction model of the state of charge of the lithium-ion battery; The preset width learning network satisfies: Wherein, Y is the prediction result output by the preset width learning network, Z n represents the mapping feature node when the current data, the voltage data, and the temperature spatiotemporal prediction data are input as input variables into the preset width learning network, Z n represents the set of all the mapping feature nodes, H m represents the enhanced node corresponding to the mapping feature node, H m represents the set of all the enhanced nodes, and are randomly generated weights and biases, ξ represents the activation function, W m is the connection weight of the width structure.

2. The method for online prediction of the state of charge of a lithium-ion battery according to claim 1, wherein: The spatiotemporal data is defined as y(x, t), and the spatiotemporal data satisfies the following relationship: In the above formula, represents the spatial basis function, a i (t) represents the low-order time coefficient, and n is the number of components that can represent the entire spatiotemporal data calculated by the KPCA algorithm; The low-order timing model is modeled according to the low-order time coefficients through the ELM model, wherein the low-order time coefficients used by the ELM model satisfy: a i (t)=f(a i (t-1),...,a i (t-d a ),u(t-1),...,u(t-d u ))+ε(t); Among them, d a and d u Represent the maximum delay of input signal and output signal respectively; The lithium-ion battery temperature prediction model satisfies: in, The temperature spatiotemporal prediction data obtained by prediction.

3. The method for online prediction of the state of charge of a lithium-ion battery according to claim 1, characterized in that: On the basis of the offline prediction model for the state of charge of the lithium-ion battery, a node is added for new input data to establish an online prediction model for the state of charge of the lithium-ion battery, and the state of charge of the lithium-ion battery is predicted by the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result, including the following sub-steps: Acquire the new input data, add nodes to the offline prediction model of the state of charge of the lithium-ion battery, and calculate the increment according to the input data; Calculating the updated weight of the node according to the incremental calculation of the input data; Applying the updated weights to the lithium-ion battery state of charge offline prediction model, and outputting the lithium-ion battery state of charge online prediction model; The state of charge of the lithium-ion battery is predicted by using the online prediction model for the state of charge of the lithium-ion battery to obtain a prediction result of the state of charge.

4. The method for online prediction of the state of charge of a lithium-ion battery as claimed in claim 3, characterized in that: Define the input data as X a , and Representing n groups of the mapping feature nodes and m groups of the enhancement nodes in the offline prediction model for the state of charge of the lithium-ion battery, the increment satisfies: in, By X a Updated incremental feature groups, and Randomly generated during the initial operation of the network; The update weights satisfy: where Y a Is additional X a The corresponding label of .

5. An online prediction system for the state of charge of a lithium-ion battery, characterized in that: include: A temperature prediction module is used to establish a lithium-ion battery temperature prediction model based on a spatiotemporal modeling strategy, and use the lithium-ion battery temperature prediction model to predict the spatiotemporal data of the lithium-ion battery temperature to obtain temperature spatiotemporal prediction data; An offline prediction module is used to establish an offline prediction model of the state of charge of the lithium-ion battery according to the spatiotemporal prediction data and the voltage data and current data of the lithium-ion battery; An online prediction module is used to add nodes for new input data based on the offline prediction model of the state of charge of the lithium-ion battery to establish an online prediction model of the state of charge of the lithium-ion battery, and predict the state of charge of the lithium-ion battery through the online prediction model of the state of charge of the lithium-ion battery to obtain a prediction result; Wherein, the temperature prediction module is also used for: Performing dimensionality reduction processing on the spatiotemporal data using a KPCA algorithm to obtain a spatial basis function that satisfies a preset number of spatial basis functions, and obtaining a corresponding low-order time coefficient; Using the ELM model to learn the low-order time coefficients, to obtain a low-order time series model with time dynamic characteristics; The subsequent time coefficients are predicted by using a low-order time series model, and then reconstructed in conjunction with the spatial basis function to obtain the lithium-ion battery temperature prediction model, and the temperature spatiotemporal prediction data is predicted by the lithium-ion battery temperature prediction model; The offline prediction module is also used for: Inputting the current data, the voltage data, and the temperature spatiotemporal prediction data as input variables into a preset width learning network for training; Outputting the trained preset width learning network as the offline prediction model of the state of charge of the lithium-ion battery; The preset width learning network satisfies: Wherein, Y is the prediction result output by the preset width learning network, Z n represents the mapping feature node when the current data, the voltage data, and the temperature spatiotemporal prediction data are input as input variables into the preset width learning network, Z n represents the set of all the mapping feature nodes, H m represents the enhanced node corresponding to the mapping feature node, H m represents the set of all the enhanced nodes, and are randomly generated weights and biases, ξ represents the activation function, W m is the connection weight of the width structure.

Images