SOC estimation method of deep fusion neural network unscented Kalman filtering

Abstract

The invention discloses an SOC (State of Charge) estimation method of deep fusion neural network unscented Kalman filtering, relates to the technical field of all-vanadium redox flow batteries, and aims to realize high-precision estimation of the state of charge of the battery. The method comprises the following steps: establishing a second-order equivalent circuit model of a battery, and performing parameter identification on the second-order equivalent circuit model; obtaining a group of sigma points through unscented transformation, calculating propagation state estimation and propagation state prediction of each sigma point, obtaining a first-order statistical moment of prior state estimation, a first-order statistical moment of prior state prediction and a cross covariance through unscented transformation, and obtaining an observation difference and a state updating difference; and inputting the observation difference and the state updating difference into a neural network system to obtain a Kalman gain, and performing state updating on the priori state estimation through the Kalman gain to obtain prediction of the state of charge of the battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of all-vanadium redox flow battery, and particularly relates to an SOC estimation method based on deep fusion neural network and unscented Kalman filtering. BACKGROUND

[0002] In recent years, with the rapid development of battery technology, the estimation of battery state of charge has attracted more and more attention. The state of charge (SOC) of the battery is one of the core parameters of the battery management system, and is an important indicator for measuring the remaining capacity or charging and discharging state of the battery.

[0003] Accurate estimation of the state of charge of the battery is beneficial to the battery management system to develop more efficient scheduling strategies, thereby preventing overcharging or overdischarging of the battery and helping to prolong the service life of the battery. However, the estimation accuracy of the state of charge of the battery is generally low at present. SUMMARY

[0004] The present application provides an SOC estimation method based on deep fusion neural network and unscented Kalman filtering, which aims to realize high-precision estimation of the state of charge of the battery.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: The present application provides an SOC estimation method based on deep fusion neural network and unscented Kalman filtering, comprising S1: establishing a second-order equivalent circuit model of the battery, and identifying the parameters of the second-order equivalent circuit model by recursive least squares; S2: obtaining a set of sigma points with fixed weights by unscented transformation, calculating the propagation state estimation and propagation state prediction of each sigma point, and obtaining the first-order statistical moments of the prior state estimation and the prior state prediction, and the cross-covariance by unscented transformation, and obtaining the observation difference and the state update difference; S3: inputting the observation difference and the state update difference into a neural network system to obtain the Kalman gain, and updating the prior state estimation by the Kalman gain to obtain the prediction of the state of charge of the battery; the neural network system comprises a long short-term memory network unit; S4: under the condition of obtaining new input variables and observation variables, updating the data of the neural network system, updating the first-order statistical moments of the prior state prediction and the cross-covariance, and cyclically executing the steps of obtaining the observation difference and the state update difference, obtaining the Kalman gain, and obtaining the prediction of the state of charge of the battery.

[0006] In some embodiments, step S1 comprises S11: under the condition of fixed battery flow, measuring the open circuit voltage of the battery by low current charging and discharging method, and obtaining the nonlinear relationship between the open circuit voltage of the battery and the state of charge of the battery. S12: establishing a second-order equivalent circuit model according to the nonlinear relationship between the open circuit voltage of the battery and the state of charge of the battery.

[0007] In some embodiments, the low current charge-discharge method measures the open circuit voltage of the battery, which refers to: when the battery is completely discharged, the battery is charged with a small current, and the voltage value and state of charge value of the battery during operation are recorded; when the battery is fully charged, the battery is discharged with a small current, and the voltage value and state of charge value of the battery during operation are recorded.

[0008] In some embodiments, each sigma point includes an initial state, The propagation state estimate and the propagation state prediction of each sigma point are calculated, which refers to: The transfer function The initial state of each sigma point is propagated to , and the predicted value is calculated by the measurement function :

[0009] The first order statistical moment of the prior state estimate and the first order statistical moment of the prior state prediction , and the cross-covariance , the observation difference and the state update difference are obtained, including:

[0010]

[0011]

[0012]

[0013] wherein, is the observation value, is the initial state value.

[0014] In some embodiments, the neural network system includes a first fully connected layer; in step S3, the observation difference and the state update difference are input into the first fully connected layer, and the output of the first fully connected layer is obtained according to ; wherein, is an activation function Relu, is a parameter of the first fully connected layer, is the output of the first fully connected layer; the parameters of the first fully connected layer include the dimension, the number of state variables and the number of observation variables.

[0015] In some embodiments, the neural network system further comprises a second fully connected layer; after obtaining the output of the first fully connected layer, step S3 further comprises: obtaining an input of the long short-term memory network unit; wherein, is a storage vector of a previous time step, is the input of the long short-term memory network unit; according to updating a state vector of the long short-term memory unit; wherein is the state vector of the previous time step, represents an updated state vector of the long-term memory unit; calculating an output vector of the storage unit , an output vector of the input gate , an output vector of the forget gate and an output vector of the output gate ; according to obtaining an output of the long short-term memory network unit ; wherein, is a Tanh activation function; according to transforming the output of the long short-term memory network unit into a Kalman gain by the second fully connected layer, wherein, is an activation function Relu, is a parameter of the second fully connected layer.

[0016] In some embodiments, the state update on the prior state estimate by the Kalman gain to obtain the prediction of the battery state of charge comprises: according to the Kalman gain and the observation difference updating a first statistical moment of the current prior state estimate ; according to obtaining a final state estimate value .

[0017] In some embodiments, the data of the neural network system is updated, the first statistical moment and the cross-covariance of the prior state prediction are updated, and the steps of obtaining the observation difference and the state update difference, obtaining the Kalman gain, and obtaining the prediction of the battery state of charge are executed in a loop, comprising: according to obtaining a covariance matrix of the battery at the next time ; wherein is the cross-covariance of the propagated state prediction.

[0018] In some embodiments, step S4 is further followed by step S5: establishing a total loss function of the second-order equivalent circuit model, and training the neural network system.

[0019] In some embodiments, the total loss function of the second-order equivalent circuit model established when training the neural network system is: .

[0020] Compared with the prior art, the application has the beneficial effects that: The SOC estimation method of the deep fusion neural network unscented Kalman filter provided by the embodiment of the application uses a second-order equivalent circuit model, which can more accurately reflect the dynamic characteristics (such as voltage delay, multiple time constants, etc.) of the battery than a first-order model, and provides a more reliable physical basis for subsequent state estimation. The recursive least squares method is used for parameter identification, which can adapt to the changes of battery parameters with aging, temperature, etc. online or offline, and maintain the accuracy of the model. The unscented transformation processes the nonlinear problem, which can more accurately capture the state distribution than the linearized extended Kalman filter (EKF), especially in high-order models, and reduces the error caused by linearization. The long short-term memory network unit can learn the complex and time-varying characteristics of the Kalman gain, which may be affected by battery aging, temperature changes, charge-discharge rate and other factors, and is difficult to accurately set by the traditional unscented Kalman filter (UKF). Through learning, the gain can be dynamically adjusted to make the state update more accurate. Therefore, a system capable of online self-adaption, high-precision and high-robustness estimation of battery SOC is formed. It not only considers the physical characteristics of the battery, but also uses data intelligence to optimize key control parameters, so as to realize more accurate measurement of the state of charge of the battery. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creating laborious work.

[0022] Figure 1 A flow chart of the SOC estimation method of the deep fusion neural network unscented Kalman filter provided by the embodiment of the application is shown in the figure. Figure 2 A schematic diagram of the structure of the all-vanadium redox flow battery provided by the embodiment of the application is shown in the figure. Figure 3 A schematic diagram of the second-order equivalent circuit provided by the embodiment of the application is shown in the figure. Figure 4 A schematic diagram of the estimation of the state of charge of the battery provided by the embodiment of the application is shown in the figure.

[0023] In the figure, 1 is a power supply, 2 is a load, 3 is a converter, 4 is a pipeline, 5 is an electrolyte tank, 6 is an ion exchange membrane, 7 is an electrode, and 8 is a circulating pump. DETAILED DESCRIPTION

[0024] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts are within the scope of the present application.

[0025] In the description of the application, it should be understood that the terms "upper", "lower", "left", "right", "front", "back", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings based on the orientation or relative position shown in the drawings, and are only for the purpose of facilitating the description of the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. Unless otherwise specified, the above directional description can be flexibly arranged in the actual application process under the condition of meeting the relative positional relationship shown in the drawings.

[0026] The terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0027] In the description of the application, it should be noted that, unless otherwise specified and limited, the terms "mounting", "connecting", "connecting", "communicating" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrally connected. It can be directly connected, or indirectly connected through an intermediate medium. It can be the communication inside two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0028] In the embodiments of the present application, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, article or device. Without more limitation, the element defined by the sentence "including a" does not exclude the presence of other identical elements in the process, article or device including the element.

[0029] In the embodiments of the present application, the words such as "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" or "for example" in the embodiments of the present application is not necessarily to be construed as preferred or advantageous over other embodiments or designs. In fact, a variety of implementations of the embodiments of the present application are possible, and the words such as "exemplary" or "for example" are simply intended to present concepts in a concrete manner.

[0030] In recent years, battery technology has developed rapidly, and electrochemical energy storage technology is an important part of it. Electrochemical energy storage technology is a technology that stores and releases electrical energy through electrochemical reactions. It converts electrical energy into chemical energy for storage and converts chemical energy back into electrical energy when needed. It is widely used in power system peak shaving and frequency regulation, renewable energy grid connection, emergency power supply, electric vehicles, and portable electronic devices.

[0031] Among various batteries, vanadium redox flow battery (VRFB) is favored due to its unique characteristics such as independent energy capacity, power output, long service life, safety, deep discharge resistance, and high efficiency. The embodiments of the present application take the all-vanadium redox flow battery as an example for illustration.

[0032] The SOC of the battery is one of the core parameters of the battery management system. With the rapid development of battery technology, the research on the estimation method of SOC has also attracted increasing attention.

[0033] The battery management system formulates the scheduling strategy based on the estimation result of the state of charge of the battery. Therefore, more accurate estimation of SOC helps the battery management system to formulate more scientific battery strategies, ensures efficient operation of the battery, prevents overcharging or overdischarging of the battery, and helps to prolong the service life of the battery.

[0034] Therefore, the embodiments of the present application provide a SOC estimation method based on deep fusion neural network unscented Kalman filtering. As shown in Figure 1 The SOC estimation method based on deep fusion neural network unscented Kalman filtering includes: S1: establishing a second-order equivalent circuit model of the battery, and identifying the parameters of the second-order equivalent circuit model by recursive least squares.

[0035] Referring to Figure 2 and Figure 3 , Figure 2 a full vanadium redox flow battery structure is shown. According to the full vanadium redox flow battery structure, a second-order equivalent circuit model as shown in Figure 3 can be established. Figure 2 The battery stack of the full vanadium redox flow battery structure in includes an ion exchange membrane 6 and an electrode 7. The electrode 7 is the site of electrochemical reaction, and the ion exchange membrane 6 can allow specific ions to pass, such as H +, which plays the role of isolating the positive and negative electrolytes. The two electrolyte storage tanks 5 are connected to the battery stack through the pipeline 4, so that the electrolyte storage tanks 5 can be used to store the positive and negative electrolytes. / and / The electrolyte can be delivered to the battery stack. A circulation pump 8 circulates the electrolyte between the electrolyte storage tank 5 and the battery stack, ensuring the reaction continues and allowing the electrolyte to continuously enter the battery stack and participate in the electrochemical reaction. The converter 3 of the all-vanadium redox flow battery connects the power source 1 and the load 2. During charging, it converts the electrical energy of the power source 1 into chemical energy for storage. During discharge, it converts the chemical energy into electrical energy for supply to the load 2, achieving bidirectional energy conversion.

[0036] More specifically, in some embodiments, step S1 also includes steps S11 and S12.

[0037] S11: When the battery flow rate is fixed, the open circuit voltage of the battery is measured using a low current charge and discharge method to obtain a nonlinear relationship between the open circuit voltage of the battery and the state of charge of the battery.

[0038] As a possible implementation method, the low-current charge and discharge method measures the open-circuit voltage of the battery, including: when the battery is completely discharged, charging the battery with a small current, and recording the voltage value and state of charge value of the battery during operation; when the battery is fully charged, discharging the battery with a small current, and recording the voltage value and state of charge value of the battery during operation.

[0039] The charge voltage and SOC curve can be plotted based on the battery's operating voltage and state of charge values ​​recorded when the battery is fully discharged and then charged with a low current. The discharge voltage and SOC curve can be plotted based on the battery's operating voltage and state of charge values ​​recorded when the battery is fully charged and then discharged with a low current. Averaging the charge voltage and SOC curve with the discharge voltage and SOC curve yields the nonlinear relationship between the battery's open circuit voltage and state of charge.

[0040] The embodiment of the present application uses the Kalman filter method to estimate the SOC of the battery. The Kalman filter is an optimal recursive filtering algorithm based on a linear dynamic system. Its core assumptions include a state space model, which includes an observation equation and a state equation. That is, establishing a second-order equivalent circuit model of the battery includes establishing a state equation and establishing an observation equation. By establishing the state equation and the observation equation, the system state can be optimally estimated through prediction and update steps. Exemplarily, the state equation and observation equation for the battery state of charge are as follows:

[0041] Among them, the state variable is ; is the sampling time, is the end voltage of the second-order RC network, and the input is the operating current , is the ohmic internal resistance, representing the instantaneous ohmic polarization of the battery, which is directly related to the electrolyte conductivity, electrode contact resistance, etc. is the electrochemical polarization branch, representing the dynamic process of charge transfer reaction; is the concentration polarization branch, representing the dynamic process of ion diffusion in the electrolyte; is the open-circuit voltage, which is a nonlinear function of SOC, ; and are independent random variables, respectively representing system noise and measurement noise; and are noise covariance matrices.

[0042] S12: Establish a second-order equivalent circuit model according to the nonlinear relationship between the open-circuit voltage of the battery and the state of charge of the battery.

[0043] After establishing the second-order equivalent circuit model, its parameters are identified by recursive least squares method. Recursive least squares (RLS) is an adaptive algorithm for parameter identification, widely used in online parameter estimation of dynamic systems. Through recursive least squares method, the key parameters in the battery model can be determined, which more accurately reflects the dynamic characteristics of the battery, thereby improving the accuracy of SOC estimation.

[0044] As a possible implementation, the least squares method is used to extract the parameters of the second-order equivalent circuit model, which requires converting the expression of the state equation into a regression form, i.e. obtaining the regression expression of the vanadium redox battery model. The regression expression of the vanadium redox battery model is obtained by using Laplace transform and bilinear transform. The obtained regression expression is as follows:

[0045] wherein, , is the information vector, is the parameter vector.

[0046] The recursive algorithm is used for updating, and the key parameters of the second-order equivalent circuit model, including resistance and capacitance, are solved by the parameter vector . Specifically, the first resistance , the second resistance , the third resistance , the first capacitance and the second capacitance .

[0047] S2: Obtain a set of sigma points with fixed weights by unscented transformation, calculate the propagation state estimation and the propagation state prediction of each sigma point, and obtain the first-order statistical moment of the prior state estimation and the first-order statistical moment of the prior state prediction, and the cross-covariance by unscented transformation, and obtain the observation difference and the state update difference.

[0048] In some embodiments, a set of sigma points is sampled for a time step t, for estimating the initial state Each sigma point represents a combination of state variables obtained from its prior distribution. Then there are:

[0049] wherein m = 3 is the number of state variables, and respectively represent the mean and covariance of the previous time step t−1. The parameter is defined as wherein = 0.8 controls the distribution of sigma points around the mean during sampling, = 0 is a secondary scaling factor.

[0050] The sigma points are obtained by sampling, and the weight of each sigma point and The calculation formula is:

[0051] wherein, = 2 is used to introduce high statistical information of state variables.

[0052] For step S2, the propagation state estimation and the propagation state prediction of each sigma point corresponding to each sigma point need to be calculated respectively according to the initial state of each sigma point. Specifically, the initial state of each sigma point is propagated to by using the transfer function , and the predicted value is calculated by using the measurement function :

[0053] As a possible implementation manner, the first-order statistical moment of the prior state estimation and the first-order statistical moment of the prior state prediction , and the cross-covariance are obtained by unscented transformation, and the observation difference and the state update difference are obtained, including:

[0054]

[0055]

[0056]

[0057] wherein, is an observation value, is a state initial value. is an initial value preset according to a second-order equivalent circuit model combined with actual conditions, is a voltage value obtained by actually measuring the battery according to a sensor or the like.

[0058] S3: inputting the observation difference and the state update difference into a neural network system to obtain a Kalman gain, and updating the prior state estimation by the Kalman gain to obtain a prediction of the battery state of charge; the neural network system comprises a long short-term memory network unit.

[0059] The neural network is an algorithm mathematical model that simulates the behavior characteristics of animal neural networks and performs distributed parallel information processing. Such a network relies on the complexity of the system to adjust the relationship between a large number of nodes connected to each other, thereby achieving the purpose of processing information.

[0060] Since the present application performs Sigma sampling near the operating point, approximates the state probability density function as a Gaussian density function of the sampling points, obtains the mean and variance of the second-order equivalent circuit model, and makes the Kalman filtering method suitable for linear assumptions also applicable to solving nonlinear system problems.

[0061] However, when the noise statistics are unknown or change over time, there is a high risk of reduced filtering accuracy or even divergence in the implementation of the unscented Kalman filter. In view of the inherent memory and nonlinear mapping capabilities of the long short-term memory neural network (also referred to as a long short-term memory network unit in the present application), it is very suitable as an internal memory element for tracking the Kalman gain. By embedding the long short-term memory unit into the unscented Kalman filtering process, noise-related information can be captured and estimated, thereby making the final battery SOC estimation result more accurate.

[0062] In some embodiments, the neural network system further comprises a first fully connected layer and a second fully connected layer. For step S3, sub-steps S31, S32, S33, S34 and S35 are included.

[0063] S31: inputting the observation difference and the state update difference into the first fully connected layer according to obtaining the output of the first fully connected layer.

[0064] wherein, is an activation function Relu, is a parameter of the first fully connected layer, is the output of the first fully connected layer; the parameter of the first fully connected layer includes dimension, number of state variables and number of observation variables.

[0065] After obtaining the output of the first fully connected layer, step S32 is performed: according to obtaining the input of the long short-term memory network unit.

[0066] wherein, is a storage vector of a previous time step, is the input of the long short-term memory network unit. according to which is iterated given any initial value, and contains high-dimensional noise-related information, is a Tanh activation function.

[0067] S33: calculating the output vector of the storage unit , the output vector of the input gate , the output vector of the forget gate and the output vector of the output gate , and updating the state vector of the long short-term memory unit according to .

[0068] wherein is a state vector of a previous time step, represents the updated state vector of the long short-term memory unit.

[0069] S34: obtaining the output of the long short-term memory network unit according to . Wherein, is a Tanh activation function.

[0070] S35: according to , the output of the long short-term memory network unit is converted into Kalman gain by the second fully connected layer.

[0071] wherein, is an activation function Relu, is a parameter of the second fully connected layer.

[0072] For the state update of the prior state estimation by the Kalman gain in step S3, the prediction of the battery state of charge specifically includes: according to the Kalman gain and the observation difference updating the first order statistics of the current prior state estimate ; according to obtaining the final state estimation value .

[0073] the final state estimation value can reflect the prediction of the battery state of charge. It should be understood that in order to be able to recursively iterate to implement the prediction of the battery state of charge at the next time, the data involved in the present application also needs to be updated: S4: updating the data of the neural network system, updating the first order statistics and cross covariance of the prior state prediction, and cyclically performing the steps of obtaining the observation difference and the state update difference, obtaining the Kalman gain, and obtaining the prediction of the battery state of charge.

[0074] Exemplarily, according to the cross covariance is updated.

[0075] Since the neural network system used in the battery state of charge estimation method provided in the present application includes a long short-term memory unit, which in turn includes an input gate, a forget gate, an output gate and a memory unit. The input gate is used to determine whether the information is stored in the state. The forget gate is used to determine whether the information needs to be discarded. The output gate determines the output information, and the memory unit is used to store long-term information. When new input variables and observation variables are input, the propagation state estimation value and the covariance matrix of the battery at the next time can be predicted and updated according to the information stored in the previous memory unit.

[0076] With reference to Figure 4 , the method for estimating the battery state of charge provided in the embodiments of the present application first samples a series of sigma points, and then passes through the transfer function , the measurement function obtains the first order statistics of the prior state estimation and the first order statistics of the prior state prediction , and the cross covariance . After calculation, the observation difference and the state update difference are obtained. Then, the Kalman gain is estimated in the neural network system to finally obtain the prediction value of SOC. The real-time update and prediction of the battery state are based on the continuous update of the state equation and the observation equation, that is, according to the continuous update of the input variable and the state variable

[0077] the continuous update of and is realized. The transfer function , measurement function and updating the first order statistics of the prior state estimate and the first order statistics of the prior state prediction and the cross covariance .

[0078] In some embodiments, step S4 is followed by step S5 of establishing an overall loss function for the second order equivalent circuit model and training the neural network system.

[0079] As one possible implementation, the overall loss function for the second order equivalent circuit model and training the neural network system is: The accuracy of the prediction is further improved through model training.

[0080] The above merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which shall be encompassed within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A SOC estimation method based on deep fusion neural network unscented Kalman filter, characterized in that: include: S1: Establishing a second-order equivalent circuit model of the battery and performing parameter identification on the second-order equivalent circuit model by recursive least squares method; S2: Obtain a set of fixed-weight sigma points through unscented transformation, calculate the propagation state estimate and propagation state prediction for each sigma point, and obtain the first-order statistical moment of the prior state estimate and the first-order statistical moment of the prior state prediction through unscented transformation, as well as the cross-covariance, to obtain the observation difference and state update difference; S3: Inputting the observation difference and the state update difference into a neural network system to obtain a Kalman gain, and performing a state update on the prior state estimate using the Kalman gain to obtain a prediction of the battery state of charge; the neural network system includes a long short-term memory network unit; S4: When new input variables and observation variable inputs are obtained, the data of the neural network system is updated, the first-order statistical moment and cross-covariance of the prior state prediction are updated, and the observation difference and state update difference are obtained in a loop, the Kalman gain is obtained, and the prediction of the battery state of charge is obtained.

2. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 1 is characterized in that: The step S1 includes: S11: When the battery flow rate is fixed, measure the open circuit voltage of the battery using a low current charge and discharge method to obtain a nonlinear relationship between the open circuit voltage of the battery and the state of charge of the battery; S12: Establishing a second-order equivalent circuit model according to the nonlinear relationship between the open circuit voltage of the battery and the state of charge of the battery.

3. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 2 is characterized in that: The low current charge and discharge method for measuring the open circuit voltage of a battery is: When the battery is completely discharged, the battery is charged with a small current, and the voltage value and the state of charge value of the battery during operation are recorded; when the battery is fully charged, the battery is discharged with a small current, and the voltage value and the state of charge value of the battery during operation are recorded.

4. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 1 is characterized in that: Each of the sigma points includes an initial state, The calculation of the propagation state estimation and propagation state prediction for each sigma point refers to: Using transfer function Propagate the initial state of each sigma point to , by measuring the function Calculate its predicted value : The first-order statistical moment of the prior state estimate is obtained by unscented transformation and the first-order statistical moment of the prior state prediction , and cross covariance , and obtain the observation difference and status update difference ,include: in, is the observed value, is the initial value of the state.

5. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 4 is characterized in that: The neural network system includes a first fully connected layer; In the step S3, The observation difference and status update difference Input the first fully connected layer, according to The output of the first fully connected layer is obtained; wherein, is the activation function Relu, are the parameters of the first fully connected layer, is the output of the first fully connected layer; the parameters of the first fully connected layer include dimension, number of state variables and number of observation variables.

6. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 5 is characterized in that: The neural network system further includes a second fully connected layer; after obtaining the output of the first fully connected layer, step S3 further includes: according to Get the input of the long short-term memory network unit; where, is the storage vector of the previous time step, The input of the long short-term memory network unit; according to Update the state vector of the long short-term memory unit; is the state vector of the previous time step, represents the updated state vector of the long-term memory unit; Calculate the output vector of the storage unit , the output vector of the input gate , the output vector of the forget gate and the output vector of the output gate ; according to Get the output of the long short-term memory network unit ;in, is the Tanh activation function; according to The output of the long short-term memory network unit Converted into Kalman gain through the second fully connected layer ,in, is the activation function Relu, are the parameters of the second fully connected layer.

7. The SOC estimation method of the deep fusion neural network unscented Kalman filter according to claim 6 is characterized in that: The step S3 further includes: According to the Kalman gain and observation difference Update the first-order statistical moment of the current prior state estimate ; according to Get the final state estimate .

8. The SOC estimation method of the deep fusion neural network unscented Kalman filter according to claim 7 is characterized in that: The step S4 comprises: according to Get the covariance matrix of the battery at the next moment ;in is the cross covariance predicted for the propagation state.

9. The SOC estimation method of a deep fusion neural network unscented Kalman filter according to claim 1 is characterized in that: After step S4, the method further includes step S5: establishing an overall loss function of the second-order equivalent circuit model and training the neural network system.

10. The SOC estimation method of the deep fusion neural network unscented Kalman filter according to claim 9 is characterized in that: The overall loss function is: 。

Images