Lithium battery state of charge estimation system and method with fast migration capability

Abstract

The present invention discloses a lithium battery state of charge estimation system and method with rapid migration capability, the system comprising: a data measurement module, a neural network training module, a network update module and a state evaluation module; the data acquisition module is used to acquire historical data of the lithium battery, the historical data comprising: historical voltage data, historical current data, historical temperature data and historical state of charge data; the neural network training module is used to construct an initial state evaluation network, and train the initial state evaluation network based on the historical data to obtain a state evaluation network; the network update module is used to measure real-time data of the lithium battery, and update the parameters of the state evaluation network based on the real-time data of the lithium battery to obtain an updated state evaluation network, the real-time data of the lithium battery comprising: real-time voltage data, real-time current data and real-time temperature data; the state evaluation module is used to perform state evaluation based on the updated state evaluation network and the real-time data of the lithium battery to obtain an evaluation result.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery state of charge estimation, and in particular relates to a lithium battery state of charge estimation system and method with fast migration capability. Background Art

[0002] Electric vehicles have attracted great interest from both industry and academia. Rechargeable lithium-ion batteries are currently the best choice for electric vehicles due to their reasonable energy / power density, cycle life, and low self-discharge rate. Although improving energy / power density can promote the adoption of electric vehicles, it may also lead to potential efficiency and safety hazards. Another key obstacle hindering the widespread adoption of electric vehicles is range anxiety. Therefore, an advanced battery management system that can provide information to monitor the battery's operating capabilities is crucial to accurately determine the state of charge and optimize battery behavior.

[0003] In essence, the state of charge is the electric vehicle’s equivalent of a fuel gauge and is defined as the ratio of remaining capacity to total capacity. It is a key parameter for multiple battery management system functions such as estimated driving range, health diagnostics, optimal charge control, balancing, etc. The state of charge cannot be measured directly. Instead, it must be estimated from measurable quantities such as terminal voltage, current, and temperature provided by the battery management system. In this context, the accuracy of the estimate and the quantification of uncertainty are very important, which enables the battery management system to make robust control and optimization decisions and avoid unexpected failures and losses. However, accurate and robust state of charge estimation remains a challenging task due to the highly nonlinear dynamic behavior, changes in temperature and load profiles, and complex degradation trajectories. Summary of the invention

[0004] The present invention aims to solve the shortcomings of the prior art and proposes a lithium battery state of charge estimation system and method with fast migration capability. For different properties of batteries at different temperatures, a lithium battery state of charge estimation model can be established using a small amount of data through fast migration learning. While estimating the lithium battery state of charge, the uncertainty interval of the estimated value is given.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A lithium battery state of charge estimation system with fast migration capability, including: a data measurement module, a neural network training module, a network update module and a state evaluation module;

[0007] The data acquisition module is used to collect historical data of the lithium battery, and the historical data includes: historical voltage data, historical current data, historical temperature data and historical state of charge data;

[0008] The neural network training module is used to construct an initial state evaluation network, and train the initial state evaluation network based on the historical data to obtain a state evaluation network;

[0009] The network update module is used to measure the real-time data of the lithium battery, and update the parameters of the state evaluation network based on the real-time data of the lithium battery to obtain an updated state evaluation network, wherein the real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data;

[0010] The state evaluation module is used to perform state evaluation based on the updated state evaluation network and the real-time data of the lithium battery to obtain an evaluation result.

[0011] Preferably, the neural network training module includes: a network construction unit and a network training unit;

[0012] The network construction unit is used to construct the initial state evaluation network, and the structure of the initial state evaluation network includes: a neural network layer consisting of a series convolutional neural network, a gated recurrent unit neural network, and a sparse Gaussian process regression layer;

[0013] The network training unit is used to train the initial state evaluation network based on the historical data and a semi-random asynchronous gradient descent algorithm to obtain the state evaluation network.

[0014] Preferably, the workflow of the network training unit includes:

[0015] Labeling the historical data to obtain an input data sequence X and corresponding state of charge data, wherein the input data sequence X includes: the historical voltage data, the historical current data, and the historical temperature data;

[0016] Initialize the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer;

[0017] Calculate a kernel matrix K based on the input data sequence X and the parameter θ;

[0018] Using the semi-stochastic asynchronous gradient descent algorithm, the parameter θ is optimized by full batch gradient descent, and the weight parameter W is optimized by random small batch gradient descent until the algorithm converges and the network training is completed;

[0019] The loss function corresponding to the gradient descent optimization is obtained by the difference between the historical charge state data and the charge state data output by the network, which is a negative logarithmic marginal likelihood function.

[0020] Preferably, the workflow of the network update module includes: using a fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain the updated state evaluation network.

[0021] The present invention also provides a lithium battery state of charge estimation method with fast migration capability, the estimation method is applied to any of the estimation systems described above, and comprises the following steps:

[0022] Collecting historical data of the lithium battery, the historical data including: historical voltage data, historical current data, historical temperature data and historical state of charge data;

[0023] Constructing an initial state evaluation network, and training the initial state evaluation network based on the historical data to obtain a state evaluation network;

[0024] Measuring real-time data of the lithium battery, and updating parameters of the state assessment network based on the real-time data of the lithium battery to obtain an updated state assessment network, wherein the real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data;

[0025] A state assessment is performed based on the updated state assessment network and the real-time data of the lithium battery to obtain an assessment result.

[0026] Preferably, the structure of the initial state evaluation network includes: a neural network layer consisting of a series convolutional neural network, a gated recurrent unit neural network, and a sparse Gaussian process regression layer.

[0027] Preferably, the training method comprises:

[0028] Labeling the historical data to obtain an input data sequence X and corresponding state of charge data, wherein the input data sequence X includes: the historical voltage data, the historical current data, and the historical temperature data;

[0029] Initialize the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer;

[0030] Calculate a kernel matrix K based on the input data sequence X and the parameter θ;

[0031] Using the semi-stochastic asynchronous gradient descent algorithm, the parameter θ is optimized by full batch gradient descent, and the weight parameter W is optimized by random small batch gradient descent until the algorithm converges and the network training is completed;

[0032] The loss function corresponding to the gradient descent optimization is obtained by the difference between the historical charge state data and the charge state data output by the network, which is a negative logarithmic marginal likelihood function.

[0033] Preferably, the updating method includes: using a fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain the updated state evaluation network.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) The present invention can effectively estimate the state of charge of battery data under different temperature conditions and provide an uncertainty interval for the result;

[0036] (2) The present invention can process long sequence data represented by battery discharge voltage, current, and surface temperature by combining convolutional neural network, gated recurrent unit neural network, and sparse Gaussian process regression, and obtain the uncertainty quantitative interval of the output result; compared with the traditional Gaussian process regression, the sparse Gaussian process regression using the kernel structure can greatly alleviate the limitation of input data length and significantly reduce the training and prediction time;

[0037] (3) The present invention has a fast transfer learning mechanism. For temperature conditions other than training data, model migration can be quickly performed by providing a small amount of training data. Compared with traditional fine-tuning transfer learning, the non-parametric transfer learning mechanism proposed in this method can significantly reduce the time spent on migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a system workflow according to an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of the working process of the network training unit according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Embodiment 1

[0045] In this embodiment, if Figure 1 , Figure 2 As shown, the lithium battery charge state estimation system with fast migration capability includes: a data measurement module, a neural network training module, a network update module and a state evaluation module

[0046] The data acquisition module is used to collect historical data of lithium batteries, including historical voltage data, historical current data, historical temperature data and historical state of charge data.

[0047] The neural network training module is used to construct an initial state evaluation network and train the initial state evaluation network based on historical data to obtain a state evaluation network. The neural network training module includes: a network construction unit and a network training unit.

[0048] The network construction unit is used to construct an initial state evaluation network. The structure of the initial state evaluation network includes: a neural network layer consisting of a series of convolutional neural networks, a gated recurrent unit neural network, and a sparse Gaussian process regression layer.

[0049] In this embodiment, the convolutional neural network is more sparse than the fully connected artificial neural network, and can efficiently extract local features and implement end-to-end estimation, that is, establish a direct mapping between input data and the battery state of charge to be estimated. The gated recurrent unit neural network is a cyclic structure with a gating mechanism. Compared with general neural networks, it can better avoid the gradient vanishing problem caused by long sequence information and reduce the possibility of training failure. Convolutional neural networks and gated recurrent unit neural networks can be regarded as combined neural network layers. Sparse Gaussian process regression uses a kernel structure approximation method to reduce the computational complexity of Gaussian process regression training, and to a certain extent, it solves the scalability problem of Gaussian process regression.

[0050] The network training unit is used to train the initial state evaluation network based on historical data and semi-random asynchronous gradient descent algorithm to obtain the state evaluation network. The workflow of the network training unit is as follows: Figure 3As shown, it includes: labeling the historical data to obtain the input data sequence X and the corresponding state of charge data, the input data sequence X includes: historical voltage data, historical current data and historical temperature data; initializing the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer; calculating the kernel matrix K based on the input data sequence X and the parameters θ; using the semi-random asynchronous gradient descent algorithm to perform full batch gradient descent optimization on the parameters θ, and performing random small batch gradient descent optimization on the weight parameters W until the algorithm converges and the network training is completed; wherein the loss function corresponding to the gradient descent optimization is obtained by the difference between the historical state of charge data and the state of charge data output by the network, which is a negative logarithmic marginal likelihood function.

[0051] Among them, sparse Gaussian process regression is realized by utilizing kernel structure, Gaussian process regression:

[0052]

[0053]

[0054]

[0055] Among them, K is the sum of the input data sequence X and the predicted input X * Determine the kernel matrix, For the effect on X * The mean function on , y is the output of the training set, and I is the identity matrix represents a normal distribution, With Cov[f * ] distribution represents the predicted noise-free output f * The expectation and covariance matrices of .

[0056] For the input of Gaussian process regression, the loss function used is the negative log marginal likelihood function NLML:

[0057]

[0058] where σ is the noise parameter of Gaussian process regression, y T represents the transpose of the training output matrix, K y represents the covariance matrix of the training output y, is the negative log marginal likelihood loss function with respect to the covariance matrix K.

[0059] In order to achieve the gradient descent of the loss function, the partial derivative of the loss function with respect to the parameter to be optimized is calculated:

[0060]

[0061] Find the partial derivative of the loss function with respect to the weight parameter W:

[0062]

[0063] Unlike the usual Gaussian process regression and neural network, the Gaussian process regression is connected in series with the neural network, so the partial derivatives of some weight parameters of the neural network cannot be directly obtained, that is, the partial derivative of K with respect to the network weight W in the formula cannot be directly obtained. Apply the chain derivation rule:

[0064]

[0065] Where H is the output of the combined neural network layer. By calculating the matrix derivative, we get:

[0066]

[0067] Through the above formula, we can calculate the gradient of a certain weight separately, that is, we can achieve factor decomposition. Next, we use the kernel structure of the Gaussian process to improve its scalability. We use the method of introducing induction points to approximate the kernel function with a regression subset:

[0068]

[0069] Where U represents the set of virtual induced points introduced, K X，X Represents the kernel function matrix of the input sequence and itself, K X，U Represents the kernel function matrix of the input sequence and the induction point sequence. The induction points can be selected arbitrarily. Here, U is selected to be evenly spaced. Due to the evenly spaced nature of the induction point set, the kernel function is further approximated:

[0070]

[0071] Among them, u j 、u a and u b is the element in the introduced induction point, u j For x i The induction point for kernel function calculation, u a 、u b For x i The two nearest induction points have u a <x i b , w i For x i to u a 、u b The linear interpolation weights are a function of the relative distance, is the constructed new approximate kernel function, converted into matrix form:

[0072] K X，U ≈WK U，U ​

[0073] Substitute the previous formula and simplify it to get:

[0074]

[0075] If the original kernel function is a static kernel function, that is, the value of the kernel function is only related to the input distance (which satisfies most kernel functions), then for one-dimensional input, it satisfies the Toeplitz structure; for multi-bit input, it satisfies the Kronecker structure. Using these structures, the computational complexity can be reduced to O(n), achieving the scalability of Gaussian process regression.

[0076] The network update module is used to measure the real-time data of the lithium battery and update the parameters of the state evaluation network based on the real-time data of the lithium battery to obtain an updated state evaluation network. The real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data.

[0077] The workflow of the network update module includes: using the fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain the updated state evaluation network.

[0078] The state assessment module is used to perform state assessment based on the updated state assessment network and real-time data of the lithium battery to obtain an assessment result.

[0079] The measured real-time voltage data, real-time current data and real-time temperature data are input into the neural network structure and then passed through the convolutional neural network to extract features. The features extracted by the convolutional neural network are input into the gated recurrent unit neural network structure, and the results are used as the input of a sparse Gaussian process regression layer. The output of the sparse Gaussian process regression layer is the battery state of charge assessment result to be obtained.

[0080] Embodiment 2

[0081] In this embodiment, a method for estimating the state of charge of a lithium battery with fast migration capability includes the following steps:

[0082] S1. Collect historical data of lithium batteries, including historical voltage data, historical current data, historical temperature data and historical state of charge data.

[0083] S2. Construct an initial state evaluation network, and train the initial state evaluation network based on historical data to obtain a state evaluation network.

[0084] The structure of the initial state evaluation network includes: a neural network layer consisting of a series of convolutional neural networks, a gated recurrent unit neural network, and a sparse Gaussian process regression layer.

[0085] In this embodiment, the convolutional neural network is more sparse than the fully connected artificial neural network, and can efficiently extract local features and implement end-to-end estimation, that is, establish a direct mapping between input data and the battery state of charge to be estimated. The gated recurrent unit neural network is a cyclic structure with a gating mechanism. Compared with general neural networks, it can better avoid the gradient vanishing problem caused by long sequence information and reduce the possibility of training failure. Convolutional neural networks and gated recurrent unit neural networks can be regarded as combined neural network layers. Sparse Gaussian process regression uses a kernel structure approximation method to reduce the computational complexity of Gaussian process regression training, and to a certain extent, it solves the scalability problem of Gaussian process regression.

[0086] The training method includes: labeling historical data to obtain an input data sequence X and corresponding state of charge data, wherein the input data sequence X includes: historical voltage data, historical current data and historical temperature data; initializing the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer; calculating the kernel matrix K based on the input data sequence X and the parameters θ; using a semi-random asynchronous gradient descent algorithm to perform full-batch gradient descent optimization on the parameters θ, and performing random small-batch gradient descent optimization on the weight parameters W until the algorithm converges and the network training is completed; wherein the loss function corresponding to the gradient descent optimization is obtained by the difference between the historical state of charge data and the state of charge data output by the network, which is a negative logarithmic marginal likelihood function.

[0087] Among them, sparse Gaussian process regression is realized by utilizing kernel structure, Gaussian process regression:

[0088]

[0089]

[0090]

[0091] Among them, K is the sum of the input data sequence X and the predicted input X * Determine the kernel matrix, For the effect on X * The mean function on , y is the output of the training set, and I is the identity matrix represents a normal distribution, With Cov[f * ] distribution represents the predicted noise-free output f * The expectation and covariance matrices of .

[0092] For the input of Gaussian process regression, the loss function used is the negative log marginal likelihood function NLML:

[0093]

[0094] where σ is the noise parameter of Gaussian process regression, y T represents the transpose of the training output matrix, K y represents the covariance matrix of the training output y, is the negative log marginal likelihood loss function with respect to the covariance matrix K.

[0095] In order to achieve the gradient descent of the loss function, the partial derivative of the loss function with respect to the parameter to be optimized is calculated:

[0096]

[0097] Find the partial derivative of the loss function with respect to the weight parameter W:

[0098]

[0099] Unlike the usual Gaussian process regression and neural network, the Gaussian process regression is connected in series with the neural network, so the partial derivatives of some weight parameters of the neural network cannot be directly obtained, that is, the partial derivative of K with respect to the network weight W in the formula cannot be directly obtained. Apply the chain derivation rule:

[0100]

[0101] Where H is the output of the combined neural network layer. By calculating the matrix derivative, we get:

[0102]

[0103] Through the above formula, we can calculate the gradient of a certain weight separately, that is, we can achieve factor decomposition. Next, we use the kernel structure of the Gaussian process to improve its scalability. We use the method of introducing induction points to approximate the kernel function with a regression subset:

[0104]

[0105] Where U represents the set of virtual induced points introduced, K X，X Represents the kernel function matrix of the input sequence and itself, K X，U Represents the kernel function matrix of the input sequence and the induction point sequence. The induction points can be selected arbitrarily. Here, U is selected to be evenly spaced. Due to the evenly spaced nature of the induction point set, the kernel function is further approximated:

[0106]

[0107] Among them, u j 、u a and u b is the element in the introduced induction point, u j For x i The induction point for kernel function calculation, ua 、u b For x i The two nearest induction points have u a <x i b , w i For x i to u a 、u b The linear interpolation weights are a function of the relative distance, is the constructed new approximate kernel function, converted into matrix form:

[0108] K X,U ≈WK U,U

[0109] Substitute the previous formula and simplify it to get:

[0110]

[0111] If the original kernel function is a static kernel function, that is, the value of the kernel function is only related to the input distance (which satisfies most kernel functions), then for one-dimensional input, it satisfies the Toeplitz structure; for multi-bit input, it satisfies the Kronecker structure. Using these structures, the computational complexity can be reduced to O(n), achieving the scalability of Gaussian process regression.

[0112] S3. Measure the real-time data of the lithium battery, and update the parameters of the state evaluation network based on the real-time data of the lithium battery to obtain an updated state evaluation network. The real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data.

[0113] The updating method includes: using a fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain an updated state evaluation network.

[0114] S4. Perform a status assessment based on the updated status assessment network and the real-time data of the lithium battery to obtain an assessment result.

[0115] The measured real-time voltage data, real-time current data and real-time temperature data are input into the neural network structure and then passed through the convolutional neural network to extract features. The features extracted by the convolutional neural network are input into the gated recurrent unit neural network structure, and the results are used as the input of a sparse Gaussian process regression layer. The output of the sparse Gaussian process regression layer is the battery state of charge assessment result to be obtained.

[0116] ​The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A lithium battery state of charge estimation system with fast migration capability, characterized in that: include: Data measurement module, neural network training module, network update module and state evaluation module; The data acquisition module is used to collect historical data of the lithium battery, and the historical data includes: historical voltage data, historical current data, historical temperature data and historical state of charge data; The neural network training module is used to construct an initial state evaluation network, and train the initial state evaluation network based on the historical data to obtain a state evaluation network; The network update module is used to measure the real-time data of the lithium battery, and update the parameters of the state evaluation network based on the real-time data of the lithium battery to obtain an updated state evaluation network, wherein the real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data; The state assessment module is used to perform state assessment based on the updated state assessment network and the real-time data of the lithium battery to obtain an assessment result; The neural network training module includes: a network construction unit and a network training unit; The network construction unit is used to construct the initial state evaluation network, and the structure of the initial state evaluation network includes: a neural network layer consisting of a series convolutional neural network, a gated recurrent unit neural network, and a sparse Gaussian process regression layer; The network training unit is used to train the initial state evaluation network based on the historical data and a semi-random asynchronous gradient descent algorithm to obtain the state evaluation network; The workflow of the network training unit includes: Labeling the historical data to obtain an input data sequence X and corresponding state of charge data, wherein the input data sequence X includes: the historical voltage data, the historical current data, and the historical temperature data; Initialize the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer; Calculate a kernel matrix K based on the input data sequence X and the parameter θ; Using the semi-stochastic asynchronous gradient descent algorithm, the parameter θ is optimized by full batch gradient descent, and the weight parameter W is optimized by random small batch gradient descent until the algorithm converges and the network training is completed; The loss function corresponding to the gradient descent optimization is obtained by the difference between the historical state-of-charge data and the state-of-charge data output by the network, which is a negative logarithmic marginal likelihood function; Among them, sparse Gaussian process regression is realized by utilizing kernel structure, Gaussian process regression: Among them, K is the sum of the input data sequence X and the predicted input X * Determine the kernel matrix, For the effect on X * The mean function on , y is the output of the training set, and I is the identity matrix represents a normal distribution, With Cov[f * ] distribution represents the predicted noise-free output f * The expectation and covariance matrices of ; For the input of Gaussian process regression, the loss function used is the negative log marginal likelihood function NLML: where σ is the noise parameter of Gaussian process regression, represents the transpose of the training output matrix, K y represents the covariance matrix of the training output y, is the negative log marginal likelihood loss function with respect to the covariance matrix K; Find the partial derivative of the loss function with respect to the parameter to be optimized: Find the partial derivative of the loss function with respect to the weight parameter W: Apply the chain derivation rule to solve the partial derivative of the covariance matrix K with respect to the network weight parameter W: Where H is the output of the combined neural network layer. By calculating the matrix derivative, we get: Use the method of introducing induced points to perform regression subset approximation on the kernel function: Where U represents the set of virtual induced points introduced, K X .X represents the kernel function matrix of the input sequence and itself, K X , U represents the kernel function matrix of the input sequence and the induced point sequence; due to the equidistant nature of the induced point set, the kernel function is further approximated: Among them, u j 、u a and u b is the element in the introduced induction point, u j For x i The induction point for kernel function calculation, u a 、u b For x i The two nearest induction points have u a <X i b ,ω i For X i to u a 、u b The linear interpolation weights are a function of the relative distance, is the constructed new approximate kernel function, converted into matrix form:​ K X，U ≈WK U，U Substitute the previous formula and simplify it to get: If the original kernel function is a static kernel function, that is, the value of the kernel function is only related to the input distance, then for one-dimensional input, it satisfies the Toeplitz structure; for multi-bit input, it satisfies the Kronecker structure; using the Toeplitz structure and the Kronecker structure, the computational complexity is reduced to O(n), achieving the scalability of Gaussian process regression; The workflow of the network update module includes: using a fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain the updated state evaluation network.

2. A lithium battery state of charge estimation method with fast migration capability, the estimation method being applied to the estimation system according to claim 1, characterized in that: The following steps are involved: Collecting historical data of the lithium battery, the historical data including: historical voltage data, historical current data, historical temperature data and historical state of charge data; Constructing an initial state evaluation network, and training the initial state evaluation network based on the historical data to obtain a state evaluation network; Measuring real-time data of the lithium battery, and updating parameters of the state assessment network based on the real-time data of the lithium battery to obtain an updated state assessment network, wherein the real-time data of the lithium battery includes: real-time voltage data, real-time current data and real-time temperature data; A state assessment is performed based on the updated state assessment network and the real-time data of the lithium battery to obtain an assessment result.

3. The method for estimating the state of charge of a lithium battery with rapid migration capability according to claim 2, characterized in that: The structure of the initial state evaluation network includes: a neural network layer consisting of a series convolutional neural network, a gated recurrent unit neural network, and a sparse Gaussian process regression layer.

4. The method for estimating the state of charge of a lithium battery with rapid migration capability according to claim 3, characterized in that: The training method comprises: Labeling the historical data to obtain an input data sequence X and corresponding state of charge data, wherein the input data sequence X includes: the historical voltage data, the historical current data, and the historical temperature data; Initialize the parameters θ of the sparse Gaussian process regression layer and the weight parameters W of the neural network layer; Calculate a kernel matrix K based on the input data sequence X and the parameter θ; Using the semi-stochastic asynchronous gradient descent algorithm, the parameter θ is optimized by full batch gradient descent, and the weight parameter W is optimized by random small batch gradient descent until the algorithm converges and the network training is completed; The loss function corresponding to the gradient descent optimization is obtained by the difference between the historical state of charge data and the state of charge data output by the network, which is a negative logarithmic marginal likelihood function.

5. The method for estimating the state of charge of a lithium battery with rapid migration capability according to claim 4, characterized in that: The updating method includes: using a fast migration mechanism to freeze the weight parameter W of the neural network layer, and performing fast migration optimization on the parameter θ of the sparse Gaussian process regression layer to obtain the updated state evaluation network.

Images