Charge state prediction method and system based on tensor network Volterra model

Abstract

The invention discloses a state-of-charge prediction method and system based on a tensor network Volterra model, and the method comprises the steps: carrying out the modeling of a lithium battery system through employing a Volterra model, collecting the nuclear coefficients of a lithium battery system model into a tensor, and carrying out the deduction of a tensor column Volterra lithium battery system model according to a specific tensor network form; training a tensor column Volterra lithium battery system model based on the discharge data, and performing optimization iteration on the model by adopting a sub-identification model-based firefly multi-innovation stochastic gradient algorithm to obtain trained model parameters; and the trained tensor column Volterra lithium battery system model is adopted to predict the state of charge of the lithium battery, and a prediction result is obtained. The unknown parameter self-adaptive online interaction estimation is realized by combining the firefly optimization algorithm and the multi-innovation stochastic gradient descent algorithm, the parameters can be identified with high precision, the calculation difficulty and complexity are remarkably reduced, and the identified model can effectively predict the state of charge of the lithium battery.

Description

Technical Field

[0001] The present invention relates to the technical field of system identification and parameter estimation, and particularly to a method and system for predicting the state of charge based on a tensor network Volterra model. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] An accurate battery model enables the BMS to effectively monitor the battery state and improve battery performance. With an accurate model, the charging and discharging of the battery can be optimized, the service life of the battery can be extended to the maximum extent, and the battery can be ensured to operate safely and reliably. The data collected by the BMS generally includes the voltage, current, and temperature of the battery, which can be directly measured by sensors. There are also some data that can be obtained through data processing and algorithm prediction methods, such as the state of charge (SOC), state of health (SOH), state of power (SOP), state of energy (SOE), and remaining useful life (RUL) of the battery. Among these data, the SOC is crucial for the energy management function of the BMS. The battery SOC reflects the remaining charge of the battery, is an instantaneous value, and reflects the current state of charge of the battery.

[0004] Establishing a mathematical model is an essential bridge to determine the connection between the actual problem at hand and mathematical tools. As the complexity of the control process increases, the application of control theory is becoming increasingly widespread, and its practical application must be based on the mathematical model of the system. However, in most cases, the mathematical model of the system is unknown. Therefore, when solving practical problems, it is necessary to first establish the mathematical model of the system, and system identification is a method of modeling, which includes system structure identification and parameter estimation. This method has been widely applied in many fields such as aerospace, bioengineering, artificial intelligence, and social economy.

[0005] System modeling and model parameter estimation are the basis of all control problems and play an important role in many practical applications. Model parameter estimation refers to determining the unknown parameters in a mathematical model so that it can best fit the observed data. Usually, some optimization algorithms are used to continuously adjust the model parameters to minimize the error between the model prediction value and the actual observed value. At present, there have been a large number of studies on linear system parameter identification methods. However, almost all actual systems contain nonlinear factors, that is, most systems are nonlinear systems. The linear system parameter identification method cannot be well applied to the analysis of nonlinear systems, which will lead to poor parameter estimation accuracy. Therefore, the research on parameter identification methods for nonlinear systems has attracted more and more attention. Compared with linear systems, nonlinear systems are more complex as a whole and have a greater variety. The Volterra series has a distinct physical meaning, can describe nonlinear behavior, and provides a powerful method and tool for solving nonlinear practical problems. Therefore, the Volterra model is widely used in constructing nonlinear system models. However, as the order increases, the Volterra kernel coefficients increase exponentially, which brings great computational complexity to directly estimating these unknown kernel coefficients. Summary of the Invention

[0006] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a state of charge prediction method and system based on a tensor network Volterra model, which combines the firefly optimization algorithm and the multi-innovation stochastic gradient descent algorithm to achieve adaptive online interactive estimation of unknown parameters, can identify parameters with high precision, significantly reduces the computational difficulty and complexity, and the identified model can effectively predict the state of charge SOC of lithium batteries.

[0007] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0008] In the first aspect, the present invention provides a state of charge prediction method based on a tensor network Volterra model, including:

[0009] Obtain the lithium battery discharge data and perform preprocessing to obtain the processed discharge data;

[0010] Use the Volterra model to model the lithium battery system, collect the kernel coefficients of the lithium battery system model into a tensor, and derive the tensor train Volterra lithium battery system model according to a specific tensor network form;

[0011] Train the tensor train Volterra lithium battery system model based on the processed discharge data, and use the firefly multi-innovation stochastic gradient algorithm based on the sub-identification model to optimize and iterate the model to obtain the trained model parameters;

[0012] The trained tensor-train Volterra lithium battery system model is used to predict the state of charge of the lithium battery, and the prediction result is obtained.

[0013] In a further technical solution, the tensor decomposition technology is adopted to represent the tensor as a product of multiple low-dimensional core tensors, and the tensor-train Volterra lithium battery system model is derived and expressed as:

[0014]

[0015] Wherein, represents a tensor, represents the transpose of the input vector, represents the TT core, and d represents the number of TT cores.

[0016] In a further technical solution, the firefly multi-innovation stochastic gradient algorithm based on the sub-identification model is specifically as follows:

[0017] The tensor-train Volterra lithium battery system model is represented as multiple sub-identification models by using the hierarchical identification principle, the loss functions of the multiple sub-identification models are respectively defined, and the loss functions are minimized by the stochastic gradient algorithm;

[0018] Combined with the multi-innovation theory, on the basis of the stochastic gradient algorithm, the training data is expanded, and the loss function is redefined to obtain the multi-innovation stochastic gradient algorithm;

[0019] A convergence exponent is introduced into the multi-innovation stochastic gradient algorithm, and the firefly optimization algorithm is used to adaptively update the step size to obtain the firefly multi-innovation stochastic gradient algorithm.

[0020] In a further technical solution, the loss function is minimized by the stochastic gradient algorithm, and the loss function is expressed as:

[0021]

[0022] Wherein, v (k) represents the vectorization of the k-th TT core, represents the expected value, y(t) represents the observed output value, b (k)T (t)v (k) represents the model output value, λ represents the regularization factor, b (k) represents the information vector.

[0023] In a further technical solution, the redefined loss function combined with the multi-innovation theory is expressed as:

[0024]

[0025] Wherein, v (k) represents the vectorization of the k-th TT core, denotes the expected value, p denotes the innovation length, Y(p, t) denotes the stacked output vector defined based on the innovation length, and Φ (k) (p, t) denotes the information matrix for the previous p time instants, and λ denotes the regularization factor.

[0026] For a further technical solution, according to the redefined loss function, p data are used each time to calculate the gradient, and the parameter v is updated at the t-th iteration (k) of the multi-innovation gradient algorithm, which is expressed as:

[0027]

[0028] where denotes the estimated value of the k-th TT core at time t, denotes the estimated value of the information matrix, denotes the estimated value of the information matrix for the previous p time instants, Y(p, t) denotes the stacked output vector defined based on the innovation length, λ denotes the regularization factor, and r k (t) denotes the step-size factor, denotes the estimated value of the information vector.

[0029] For a further technical solution, when the firefly optimization algorithm reaches the maximum number of iterations, the update stops, and the optimal solution is the value of the convergence exponent, obtaining the firefly multi-innovation stochastic gradient algorithm, which is expressed as:

[0030]

[0031] Y(p, t) = [y(t), y(t - 1),..., y(t - p + 1)] T

[0032]

[0033] where denotes the estimated value of the k-th TT core at time t, denotes the estimated value of the information matrix, denotes the estimated value of the information matrix for the previous p time instants, Y(p, t) denotes the stacked output vector defined based on the innovation length, λ denotes the regularization factor, and r k (t) denotes the step-size factor, denotes the estimated value of the information vector, denotes the estimated value of the convergence exponent, and FA[] denotes the firefly optimization algorithm.

[0034] In a second aspect, the present invention provides a state-of-charge prediction system based on a tensor network Volterra model, including:

[0035] A data acquisition module, which is configured to: acquire lithium battery discharge data and perform preprocessing to obtain processed discharge data;

[0036] A model establishment module, which is configured to: model the lithium battery system using the Volterra model, collect the kernel coefficients of the lithium battery system model into a tensor, and derive a tensor train Volterra lithium battery system model according to a specific tensor network form;

[0037] A model training module, which is configured to: train the tensor train Volterra lithium battery system model based on the processed discharge data, and optimize and iterate the model using a firefly multi-innovation stochastic gradient algorithm based on a sub-identification model to obtain trained model parameters;

[0038] A model prediction module, which is configured to: predict the state of charge of the lithium battery using the trained tensor train Volterra lithium battery system model to obtain a prediction result.

[0039] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps in the state of charge prediction method based on the tensor network Volterra model described in the first aspect are implemented.

[0040] In a fourth aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in the state of charge prediction method based on the tensor network Volterra model described in the first aspect are implemented.

[0041] The above one or more technical solutions have the following beneficial effects:

[0042] The present invention optimizes the gradient algorithm in combination with the multi-innovation theory. By establishing a data window of length p, not only the current data is used in each iteration, but also the information of the past p time instants is introduced, increasing the data utilization rate, reusing the system data, improving the convergence speed of the algorithm, reducing the estimation error, using the firefly optimization algorithm to dynamically update the step size, and realizing the real-time estimation of the model parameters through repeated optimization and iteration, and then completing the output prediction of the test data set, so that the model can better fit the observed data.

[0043] The present invention stores the kernel coefficients in the Volterra model in tensor form and uses tensor decomposition technology to represent the Volterra kernel, obtaining a reduced-dimensional representation of the Volterra model. Compared with the traditional Volterra model, it significantly reduces the parameters, significantly reducing the computational difficulty and complexity. Description of the Drawings

[0044] The accompanying drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0045] Figure 1 It is the overall flowchart of the model parameter optimization in the embodiment of the present invention;

[0046] Figure 2 It is the graph of the change of the model training error in the embodiment of the present invention under different innovation lengths;

[0047] Figure 3 It is the graph of the change of the model training error in the embodiment of the present invention under different signal-to-noise ratios. Detailed implementation manners

[0048] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0049] It should be noted that the terms used herein are only for describing the specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0050] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0051] To solve the problem that the Volterra kernel coefficients increase exponentially, which brings great computational complexity to directly estimate these unknown kernel coefficients, the present invention stores the kernel coefficients in tensor form and uses tensor decomposition technology to represent the Volterra kernel, obtaining a reduced-dimensional representation of the Volterra model. Compared with the traditional Volterra model, it significantly reduces the parameters. Since the prediction of the state of charge (SOC) of lithium batteries has high requirements for real-time performance, it is very necessary to develop an efficient online identification method to solve this problem. The present invention proposes an adaptive online parameter identification method combining the firefly optimization algorithm and the multi-innovation stochastic gradient descent algorithm, which can identify parameters with high accuracy and significantly reduce the computational difficulty and complexity.

[0052] Embodiment 1

[0053] As Figure 1As shown, this embodiment discloses a method for predicting the state of charge based on a tensor network Volterra model, and the method includes the following steps:

[0054] S1: Obtain the discharge data of the lithium battery and perform preprocessing to obtain the processed discharge data;

[0055] In this embodiment, the discharge data of the lithium battery is obtained to form a sample set, which includes the current and voltage data obtained at a certain sampling interval during the discharge process, namely the input samples u1,..., u l , and the state of charge (SOC) data of the battery, namely the output samples y1,..., y l , is calculated through the current and voltage data. To eliminate the adverse effects caused by singular sample data, the data is normalized and divided into a training data set and a test data set.

[0056] S2: Use the Volterra model to model the lithium battery system, collect the kernel coefficients of the lithium battery system model into a tensor, and derive the tensor train Volterra lithium battery system model according to a specific tensor network form;

[0057] In this embodiment, the lithium battery system can be characterized by the Volterra model. Considering a d-order Volterra system with l inputs and a single output, all kernel coefficients are collected into a Volterra tensor, and the TT-MISO-Volterra model (multi-input single-output tensor train Volterra model) is derived according to a specific tensor network form. The rank of the TT kernel is set to Initialize the parameters: step size factor r, memory delay M, TT kernel, regularization factor λ.

[0058] Establish a multi-input single-output Volterra model (MISO-Volterra), which specifically includes the following steps:

[0059] S201: A d-order Volterra system with l inputs and a single output can be expressed as:

[0060]

[0061] where M is the memory delay, i k represents the delay index at different orders k, h k represents the kernel coefficient, k represents the current non-linear order, n represents the quadrature operation index at the current order k, and u(t - i n ) represents the value of the input vector at the past i k moments.

[0062] Since the Volterra kernel coefficients grow exponentially with the increase of the order, all kernel coefficients h k (i1, ..., i k ) are collected into a tensor with the dimension of (lM + 1) × … × (lN + 1). Multiple inputs are extensions based on a single input. Integrating multiple inputs into an input vector u t gives:

[0063]

[0064] where

[0065] u t = [1, u1(t), u2(t), ..., u l (t), ..., u1(t - M + 1), u2(t - M + 1), …, u1(t - M + 1)] T

[0066]

[0067] where denotes the Kronecker product, and T denotes the transpose operation.

[0068] S202: According to the preset TT-rank, the Volterra tensor is stored in the TT format using the tensor decomposition technique (Tensor-Train decomposition, also known as tensor train decomposition, abbreviated as TT decomposition). The original Volterra high-dimensional tensor is represented as the product of d low-dimensional kernel tensors, and the formula is expressed as:

[0069]

[0070] where i d represents the size of the d-th dimension.

[0071] The Volterra tensor has its storage requirement reduced to O(d(lM + 1)r 2 ), and the expression of the TT-MISO-Volterra model (multi-input single-output tensor train Volterra model) can be obtained as follows:

[0072]

[0073] where denotes a tensor, u t denotes the input vector, denotes the transpose of the input vector,

[0074] Denote the low - dimensional core tensor (also known as the TT - core), \(d\) represents the number of low - dimensional core tensors, and the first TT - core has dimensions \((LM + 1)\times r_1\), and the \(k\) - th TT - core has dimensions \(r\) k-1 \(\times(lM + 1)\times r\) k , and the last TT - core has dimensions \(r\) d-1 \(\times(lM + 1)\).

[0075] The TT - core is decomposed from the high - dimensional Volterra core tensor. Through TT - decomposition, the high - dimensional Volterra core tensor is decomposed into the product of multiple low - dimensional core tensors. The low - dimensional core tensors obtained here are the TT - cores. The core idea of hierarchical identification is to decompose the system into multiple sub - identification models, and these TT - cores are the unknown parameters in each sub - identification model.

[0076] S3: Train the Tensor - Train Volterra lithium - battery system model based on the processed discharge data, and optimize and iterate the model using the firefly multi - innovation stochastic gradient algorithm based on the sub - identification model to obtain the trained model parameters;

[0077] In this embodiment, the firefly multi - innovation stochastic gradient algorithm based on the sub - identification model is specifically as follows:

[0078] S301: Use the hierarchical identification principle to represent the Tensor - Train Volterra lithium - battery system model as \(d\) sub - identification models. The sub - identification models are pseudo - linear regression models regarding each TT - core; Based on the \(d\) sub - identification models, define their loss functions respectively, and minimize the loss functions through the stochastic gradient algorithm;

[0079] Construct the pseudo - linear regression model regarding the \(k\) - th TT - core (i.e., the \(k\) - th sub - identification model) as follows:

[0080]

[0081] where, represents the transpose of the input vector, represents the \(k\) - th TT - core, \(vec()\) is an operator that represents converting a matrix into a vector, \(v\) (k) is the vectorization of the \(k\) - th TT - core and is also the parameter vector to be estimated; is the information vector.

[0082] The loss function is shown as the following formula:

[0083]

[0084] where, \(v\) (k) represents the vectorization of the \(k\) - th TT - core, denotes the expected value, y(t) denotes the observed output value, and b (k)T (t)v (k) denotes the model output value, λ denotes the regularization factor, and b (k) denotes the information vector.

[0085] Using the negative gradient search principle to optimize the loss function to make it reach the minimum value, the estimated parameter v at the t-th iteration can be obtained (k) The optimization algorithm (stochastic gradient descent algorithm) is as follows:

[0086]

[0087] where denotes the estimated value of the k-th TT kernel, denotes the gradient, denotes the estimated value of the information vector, and r k (t) denotes the step size factor.

[0088] S302: Combining the multi-innovation theory, on the basis of the stochastic gradient algorithm, expand the training data, redefine the loss function, and obtain the multi-innovation stochastic gradient algorithm; specifically, combining the multi-innovation theory, consider the influence of the data in the previous period on the parameter update at the current moment, that is, expand the data used in each training, and use the p sample data in the past period of time to calculate the gradient and update the parameters.

[0089] According to the multi-innovation theory, expand the training data and apply it to model training. The update steps of the parameters to be identified (unknown parameters in the sub-identification model) specifically include:

[0090] (1) Define the innovation at time t as Considering the influence of the training data in the previous period on the parameter update at the current moment, add the data in the past p moments to expand the innovation, improve the data utilization rate, and reuse the system data. Since the parameter estimated value at time t - 1 is closer to the true value than the parameter estimated values at the previous moments t - i (i = 2, 3,..., p - 1) Therefore, for simplicity, the expanded innovation vector can be reasonably expressed as:

[0091]

[0092] where p is called the innovation length, and E(p, t) denotes the innovation vector. The parameters are iteratively updated at each moment, and the parameter update at time t also represents the t-th iteration of the algorithm.

[0093] (2) Define the stacked output vector Y(p, t) and the expanded information matrix Φ(p, t) based on the innovation length p as:

[0094] Y(p, t) = [y(t), y(t - 1),..., y(t - p + 1)] T

[0095] Φ(p, t = [b (k) (t), b (k) (t - 1),..., b (k) (t - p + 1)] T 。

[0096] (3) Redefine the loss function in step S301 as follows:

[0097]

[0098] where v (k) is the vectorization of the k-th TT core, denotes the expected value, Y(p, t) is the stacked output vector defined based on the innovation length, Φ(p, t) is the extended information matrix, and λ is the regularization factor.

[0099] According to the above loss function, each time p sample data are used to calculate the gradient, the multi-innovation stochastic gradient algorithm for updating the parameter v (k) at the t-th iteration can be obtained as follows:

[0100]

[0101] where, denotes the estimated value of the k-th TT core (i.e., the vector of parameters to be estimated in the k-th sub-identification model) at time t, denotes the estimated value of the information matrix, denotes the estimated value of the information matrix at the previous p time instants, Y(p, t) denotes the stacked output vector defined based on the innovation length, Φ(p, t) denotes the extended information matrix, λ denotes the regularization factor, r k (t) denotes the step size factor, denotes the estimated value of the information vector.

[0102] S303: Introduce the convergence exponent ε into the multi-innovation stochastic gradient algorithm, and adopt the Firefly Algorithm to adaptively update the step size to obtain the Firefly multi-innovation stochastic gradient algorithm. Use the training set data to perform online training on multiple sub-identification models, where each iteration updates the parameters according to the current observed values and the parameter estimates obtained in the previous iteration, and gradually optimizes the parameter estimates to adapt to the observed data. During the training process of the model, continue to iterate until the preset number of iterations is reached or the stop condition is satisfied. Use the root mean square error RMSE as the index to evaluate the performance of the model.

[0103] The specific steps for repeatedly training and optimizing the model using the firefly multi-innovation stochastic gradient algorithm are as follows:

[0104] (1) Set the initial parameters of the firefly optimization algorithm, including the firefly population size n, the light absorption coefficient γ of the medium, the dimension d, the initial step size coefficient α, the maximum attractiveness β max , the minimum attractiveness β min , the upper and lower limit value range h of the population, and the maximum number of iterations T.

[0105] (2) Randomly generate n firefly individuals within the value range h, calculate the fitness function value f of each firefly individual in the firefly population. The brighter the firefly with a better fitness value, retain the optimal fitness function value f best , and take the optimal firefly individual p best corresponding to this fitness function as the current global optimal solution. The fitness function value f is expressed as:

[0106]

[0107] (3) Update the positions of the firefly individuals. For the i-th non-optimal individual p i , each firefly will fly towards a firefly with a higher brightness than itself, and its position update is shown as follows:

[0108]

[0109]

[0110] α(t) = α′

[0111] where p′ i represents the position of the firefly individual after update, β represents the attractiveness, p j represents the position of a firefly with a higher brightness than the i-th individual p i , γ represents the light absorption coefficient of the medium, r represents the distance between the i-th firefly and the j-th firefly, rand() is a random perturbation between [-0.5, 0.5], and the firefly flight step size α will decrease with time. α t represents the flight step size of the firefly at the t-th generation, and α is the step size factor of the perturbation.

[0112] (4) Use the perturbation mechanism to update the optimal individual p best , that is, based on the position of the current optimal individual, add a random perturbation term to obtain a new individual. The update formula is as follows:

[0113] p best ′ = p best + αrandGuass(()

[0114] Among them, p′ best is the updated optimal individual, and Guass() represents sampling random numbers from a Gaussian distribution.

[0115] (5) When the algorithm reaches the maximum number of iterations, the update stops, and the positions of the fireflies searched are used as the solution output. That is, the obtained optimal solution is the value of the convergence exponent ε, and the firefly multi-new information stochastic gradient algorithm based on the sub-identification model is obtained:

[0116]

[0117]

[0118]

[0119] Y(p, t) = [y(t), y(t - 1),..., y(t - p + 1)] T

[0120]

[0121] Among them, is the estimated value of the convergence exponent, and FA[] represents the firefly optimization algorithm.

[0122] The unknown parameter adaptive online interactive estimation is realized by successively estimating the unknown parameters in the sub-identification model through the firefly multi-new information stochastic gradient algorithm, and the parameters can be identified with high precision.

[0123] (6) Using all the training data, optimize and iterate the model according to the firefly multi-new information stochastic gradient algorithm based on the sub-identification model, update each TT kernel in turn, and stop when the maximum number of iterations is reached. The finally trained model parameters can be used for subsequent prediction of the state of charge (SOC) of lithium batteries.

[0124] The index for evaluating the model performance is the root mean square error RMSE, as shown in the following formula:

[0125]

[0126] Among them, y(t) represents the true output corresponding to the input sample at time t, represents the model output corresponding to the input sample at time t, and N represents the total number of training samples.

[0127] S4: Use the trained tensor train Volterra lithium battery system model to predict the state of charge of the lithium battery and obtain the prediction result.

[0128] In this embodiment, the model parameters after training are saved, the predicted output is calculated through the input sample data in the test dataset, and compared with the true output to calculate the error, verifying the effectiveness of the algorithm.

[0129] To verify the effectiveness of the proposed solution of the present invention, a simulation experiment is carried out below.

[0130] Consider the following 3rd-order TT-MISO-Volterra model:

[0131]

[0132] Among them, represents the TT kernel, and v(t) represents the noise.

[0133] Combined with the hierarchical identification principle, the model is decomposed into sub-identification models based on 3 TT kernels:

[0134]

[0135]

[0136]

[0137]

[0138]

[0139] Among them,

[0140] The data for this experiment is randomly generated by the system, that is, by randomly generating the input, obtaining the system output, using the generated training sample data to optimize the model parameters, and then predicting the system output through the test sample data. The root mean square error is used as the index to evaluate the model performance. Algorithm parameter setting: the sample data is 3000, the training data is 2900, and the test data is 100. In the experiment, the performance of the algorithm is verified by using different innovation lengths p and different signal-to-noise ratios. Figure 2 shows the root mean square error change curve under different p values. As the number of iterations increases, the error decreases, and at the same number of iterations, the optimization algorithm based on multi-innovation performs better. From Figure 2 it can be seen that as the innovation length increases, the convergence speed accelerates, the error decreases significantly, and the parameter estimation accuracy is higher. Figure 3 shows the root mean square error change curve under different signal-to-noise ratios SNR. As the signal-to-noise ratio increases, the error change tends to be flat and decreases, that is, the lower the noise level, the better the estimation effect of the proposed method. Therefore, the simulation results prove that the proposed algorithm is effective for the identification of the TT-MISO-Volterra model.

[0141] Example 2

[0142] This embodiment discloses a state of charge prediction system based on a tensor network Volterra model, including:

[0143] A data acquisition module, which is configured to: acquire lithium battery discharge data and perform preprocessing to obtain processed discharge data;

[0144] A model establishment module, which is configured to: use the Volterra model to model the lithium battery system, collect the kernel coefficients of the lithium battery system model into a tensor, and derive a tensor train Volterra lithium battery system model according to a specific tensor network form;

[0145] A model training module, which is configured to: train the tensor train Volterra lithium battery system model based on the processed discharge data, and optimize and iterate the model using a firefly multi-innovation stochastic gradient algorithm based on a sub-identification model to obtain trained model parameters;

[0146] A model prediction module, which is configured to: predict the state of charge of the lithium battery using the trained tensor train Volterra lithium battery system model to obtain a prediction result.

[0147] Example 3

[0148] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method in Example 1 are implemented.

[0149] Example 4

[0150] The purpose of this embodiment is to provide a computer-readable storage medium. A computer-readable storage medium stores a computer program, and when the program is executed by a processor, the steps of the method in Example 1 are executed.

[0151] The steps involved in the devices in the above Examples 3 and 4 correspond to those in Method Example 1, and the specific implementation details can be referred to the relevant description part of Example 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0152] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by the computing device, so that they can be stored in the storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0153] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0154] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A method for predicting state of charge based on a tensor network Volterra model, characterized in that: include: Acquire lithium battery discharge data and perform preprocessing to obtain processed discharge data; The Volterra model is used to model the lithium battery system, the kernel coefficients of the lithium battery system model are collected into a tensor, and the tensor sequence Volterra lithium battery system model is derived according to a specific tensor network form; The tensor sequence Volterra lithium battery system model is trained based on the processed discharge data, and the model is optimized and iterated using the Firefly multi-innovation stochastic gradient algorithm based on the sub-identification model to obtain trained model parameters; The trained tensor sequence Volterra lithium battery system model is used to predict the state of charge of the lithium battery and obtain the prediction results.

2. The method for predicting state of charge based on the tensor network Volterra model according to claim 1, characterized in that: By using tensor decomposition technology, the tensor is expressed as the product of multiple low-dimensional core tensors, and the tensor sequence Volterra lithium battery system model is derived, which is expressed as: in, represents a tensor, represents the transpose of the input vector, represents the TT core, and d represents the number of TT cores.

3. The method for predicting state of charge based on the tensor network Volterra model according to claim 1, characterized in that: The firefly multi-innovation stochastic gradient algorithm based on the sub-identification model is specifically: The tensor sequence Volterra lithium battery system model is represented as multiple sub-identification models by using the hierarchical identification principle, and the loss functions of the multiple sub-identification models are defined respectively, and the loss functions are minimized by the stochastic gradient algorithm. Combining the multi-innovation theory and based on the stochastic gradient algorithm, the training data is expanded and the loss function is redefined to obtain the multi-innovation stochastic gradient algorithm. The convergence index is introduced into the multi-innovation stochastic gradient algorithm, and the firefly optimization algorithm is used to adaptively update the step size to obtain the firefly multi-innovation stochastic gradient algorithm.

4. The method for predicting state of charge based on the tensor network Volterra model according to claim 3, characterized in that: The loss function is minimized by the stochastic gradient algorithm, which is expressed as: Among them, v (k) represents the vectorization of the kth TT kernel, represents the expected value, y(t) represents the observed output value, and b (k)T (t)v (k) represents the model output value, λ represents the regularization factor, b (k) represents the information vector.

5. The method for predicting state of charge based on the tensor network Volterra model according to claim 3, characterized in that: Combined with the multi-innovation theory, the redefined loss function is expressed as: Among them, v (k) represents the vectorization of the kth TT kernel, represents the expected value, p represents the innovation length, Y(p, t) represents the stacked output vector defined based on the innovation length, Φ (k) (p, t) represents the information matrix of the previous p moments, and λ represents the regularization factor.

6. The method for predicting state of charge based on the tensor network Volterra model according to claim 5, characterized in that: According to the redefined loss function, p data are used to calculate the gradient each time to get the tth iteration update parameter v (k) The multi-innovation gradient algorithm is expressed as: in, represents the estimated value of the kth TT core at time t, represents the estimated value of the information matrix, represents the estimated value of the information matrix of the previous p moments, Y(p,t) represents the stacked output vector defined based on the length of the new information, λ represents the regularization factor, r k (t) represents the step size factor, Represents the estimated value of the information vector.

7. The method for predicting state of charge based on the tensor network Volterra model according to claim 3, characterized in that: When the firefly optimization algorithm reaches the maximum number of iterations, it stops updating and outputs the optimal solution as the value of the convergence index. The firefly multi-innovation stochastic gradient algorithm is obtained, which is expressed as: Y(p,t)-[y(t),y(t-1),...,y(t-p+1)] T in, represents the estimated value of the kth TT core at time t, represents the estimated value of the information matrix, represents the estimated value of the information matrix at the previous p moments, Y(p, t) represents the stacked output vector defined based on the length of the new information, λ represents the regularization factor, r k (t) represents the step size factor, represents the estimated value of the information vector, represents the estimated value of the convergence index, and FA[] represents the firefly optimization algorithm.

8. A state of charge prediction system based on a tensor network Volterra model, characterized in that: include: A data acquisition module is configured to: acquire and pre-process the lithium battery discharge data to obtain processed discharge data; A model building module is configured to: use the Volterra model to model the lithium battery system, collect the kernel coefficients of the lithium battery system model into a tensor, and derive the tensor sequence Volterra lithium battery system model according to a specific tensor network form; A model training module is configured to: train the tensor sequence Volterra lithium battery system model based on the processed discharge data, and optimize the model iteratively using a firefly multi-innovation stochastic gradient algorithm based on a sub-identification model to obtain trained model parameters; The model prediction module is configured to: use the trained tensor sequence Volterra lithium battery system model to predict the state of charge of the lithium battery to obtain a prediction result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the charge state prediction method based on the tensor network Volterra model as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the charge state prediction method based on the tensor network Volterra model as described in any one of claims 1-7 are implemented.

Images