A Mamba architecture-driven battery pack SOC and SOE estimation method

Abstract

The present invention discloses a method for estimating the SOC and SOE of a battery pack driven by a Mamba architecture. The method first collects and preprocesses the original data of an electric vehicle battery pack, and makes the preprocessed battery data into a data set. Secondly, a Mamba network is constructed, and the Mamba network model includes a fully connected layer, a feature layer regularization, a gating unit, and a selective state space model SSM module. Finally, two task layers are built, and the output of the Mamba network is used to respectively estimate the SOC and SOE of the electric vehicle battery pack. The present invention estimates the SOC and SOE of the battery pack simultaneously, improves the calculation efficiency of the model, reduces the calculation and storage requirements, has a simple measurement process, and can timely and accurately estimate the SOC and SOE of the lithium-ion battery.

Description

Technical Field

[0001] The present invention relates to the technical field of power battery management, and in particular to a method for jointly estimating SOC and SOE of an electric vehicle battery pack driven by a Mamba architecture. Background Art

[0002] The research on battery SOC (state of charge) and SOE (state of energy) estimation can be divided into two categories: model-based methods and data-driven methods. Model-based methods reflect the internal degradation process of the battery by establishing a physical model. However, due to the complexity of the internal reaction of the battery and the great influence of external conditions, it is difficult to establish an accurate model. On the other hand, the data-driven method ignores the internal mechanism and directly establishes a mapping relationship with the battery capacity through historical data. This method is divided into traditional machine learning methods and artificial neural network methods. Artificial neural network methods have attracted much attention from researchers due to their good generalization performance and strong feature extraction capabilities. Among them, recurrent neural networks are a type of neural network specifically used to process time series information, and are also one of the most commonly used neural networks for estimating SOC and SOE. However, recurrent neural networks are limited in their ability to process fairly long sequence information and cannot fully process the previous and next correlation information. And for the two parameters of SOC and SOE, existing research focuses on the estimation of one parameter, which undoubtedly increases the computational complexity. Summary of the invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and propose a Mamba architecture-driven battery pack SOC and SOE estimation method that allows the model to selectively transmit or forget information based on current data. This method not only solves the shortcomings of traditional networks in processing discrete intensive data, but also has extremely fast computing speed and has the potential to be applied to electric vehicle battery management systems. It can simultaneously obtain high-precision joint estimation results of battery pack SOC and SOE.

[0004] In order to solve the above technical problems, the technical solution of the present invention is:

[0005] A Mamba architecture driven battery pack SOC and SOE estimation method comprises the following steps:

[0006] S1. Collect and preprocess the original data of the electric vehicle battery pack, and make the preprocessed battery data into a data set to train and test the network model. The original data of the battery pack includes the real-time running mileage, speed, external temperature, operating temperature, maximum voltage, minimum voltage, total voltage, current and corresponding SOC and SOE data of the battery pack.

[0007] S2. Construct a Mamba network model, wherein the Mamba network includes a fully connected layer, a feature layer regularization, a gating unit and a selective state space model SSM module.

[0008] Joint estimation of SOC and SOE of electric vehicle battery pack driven by S3 and Mamba architecture. Two task layers are built to respectively estimate the SOC and SOE of electric vehicle battery pack. For both tasks, the highest voltage-time of the battery pack, the lowest voltage-time of the battery pack, and the real-time total working voltage-time of the battery pack in the data set are used as network inputs. The SOC and SOE results integrated in the corresponding vehicle-side battery management system are used as labels, and the estimation errors of SOC and SOE are calculated.

[0009] Preferably, in step S1, the electric vehicle battery pack raw data collection method is as follows: the real-time operation, charging and static data of the electric vehicle are uploaded to the cloud platform, and the real-time operation mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, battery pack real-time total operating voltage, battery pack current and corresponding battery pack SOC and battery pack SOE data of the electric vehicle battery pack are obtained from the cloud platform.

[0010] Preferably, the electric vehicle battery pack raw data preprocessing method in step S1 is:

[0011] First, all the original data are cleaned, that is, the abnormal points are deleted. Then, the Pearson correlation coefficient method is used to perform correlation analysis on the real-time running mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, battery pack real-time total operating voltage, battery pack current, battery pack SOC, and battery pack SOE of the electric vehicle battery pack, and the three groups of original data with the greatest correlation with the battery pack SOC and SOE are obtained. In the present invention, the battery pack maximum voltage, battery pack minimum voltage, and battery pack real-time total operating voltage are selected as the original input data.

[0012] Then, the original input data is normalized by the min-max method, and the processed data is recorded as x input =[V max ,V min ,V pack ], where V max 、V min 、V pack They represent the maximum voltage, minimum voltage and real-time total working voltage of the electric vehicle battery pack respectively.

[0013] Finally, a sliding time window is set for the input three-dimensional data, and the mathematical expression of the input data is as follows:

[0014]

[0015] t represents the tth timestamp, and w represents the sliding window size.

[0016] Preferably, in step S1, the preprocessed data is made into a data set, and the data set is divided into a training set, a validation set and a test set in a ratio of 5:1:1.

[0017] Preferably, in step S2, the Mamba network architecture construction method is:

[0018] First, the input data is regularized through the feature layer to obtain Input two fully connected layers in two ways and get and

[0019] Secondly, the gating unit includes a SiLU activation function and a multiplication operation, After the depth-separable convolutional layer, the SiLU activation function is input and the result is After SiLU activation function, we get

[0020] Then, After the selective SSM module, it enters the feature layer regularization to obtain and Perform multiplication to get

[0021] final, After the fully connected layer Add together to get the final output X O .

[0022] Preferably, the step S2 includes the following sub-steps:

[0023] Optional SSM module construction: First, build the state space model (SSM) module, which captures the system state variables by using a set of first-order differential equations:

[0024]

[0025] y(t)=Ch(t)+Dx(t)

[0026] in, Represents the derivative of the state vector h(t), and matrices A, B, C, and D define the relationship between h(t), x(t), and y(t). Then, since SSM cannot process discrete data, SSM is adapted to discrete data:

[0027]

[0028]

[0029] This discrete form of SSM can effectively establish the hidden state h k Related recursive relations. In the above formula, They are composed of parameter matrices. Usually used as a residual connection, it can be simplified in some cases. From continuous time signal to discrete time signal, a step size Δ is needed to transition, thereby converting the continuous signal x(t) into a discrete signal x k =x(kΔ). Thus, the following derivation formula is obtained:

[0030]

[0031]

[0032]

[0033] in,(·) -1 Represents the inverse of the matrix. For SSM, it has the inherent linear time invariant property, and the same matrix is ​​used for different inputs. This limits the ability to understand the differences between different sequences. Therefore, a selective SSM module is built. The selective SSM module will and Δ are set to dynamic elements that change according to the input, so that the selective SSM module can effectively adapt to different input context information. Specifically, the fully connected layer is embedded in the selective SSM module. and Δ for adjustability learning.

[0034] Preferably, in step S3, the task layer construction method is:

[0035] After the feature passes through the Mamba network, feature X is obtained mamba After the Mamba module, two task layers are serially connected. The task layers are composed of a fully connected layer, a softmax activation function, a flatten layer, a ReLU activation function, a dropout layer, and a fully connected layer in series, and the final two prediction outputs are obtained, namely the estimated SOC and SOE of the electric vehicle battery pack at each moment.

[0036] The present invention has the following characteristics and beneficial effects:

[0037] By adopting the above technical solution, the present invention uses the Mamba network architecture to process the raw data with the greatest correlation with SOC and SOE in the real-time operation data of electric vehicles, and obtains the dynamic information of the battery pack of electric vehicles by introducing a selective mechanism to improve the model's understanding and estimation ability of time series data. Through the dual-task estimation method driven by the Mamba architecture, the battery pack SOC and SOE can be estimated simultaneously, improving the model calculation efficiency and reducing the calculation and storage requirements. The measurement process is simple, and the SOC and SOE of lithium-ion batteries can be estimated in a timely and accurate manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 This is a framework diagram of a method for jointly estimating SOC and SOE of an electric vehicle battery pack driven by a Mamba architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention provides a Mamba architecture driven battery pack SOC and SOE estimation method, such as Figure 1 As shown, the following steps are included:

[0041] Step (1): Electric vehicle battery pack raw data collection and preprocessing.

[0042] (a) All raw data of electric vehicle battery packs are cleaned, that is, abnormal points are deleted, including the real-time running mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, battery pack real-time total operating voltage, battery pack current and corresponding battery pack SOC and battery pack SOE data of electric vehicle battery packs.

[0043] (b) The Pearson correlation coefficient method is used to perform correlation analysis on the real-time running mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, battery pack real-time total operating voltage, battery pack current, battery pack SOC, and battery pack SOE of the electric vehicle battery pack, and the three groups of original data with the greatest correlation with the battery pack SOC and SOE are obtained. The mathematical expression of the Pearson correlation coefficient method is as follows:

[0044]

[0045] Among them, x i and iThey represent the real-time running mileage, speed, external temperature, operating temperature, maximum voltage, minimum voltage, total voltage, current, SOC and SOE of the battery pack of the electric vehicle. and In this invention, we select the maximum voltage of the battery pack, the minimum voltage of the battery pack, and the real-time total working voltage of the battery pack as the original data.

[0046] (c) The data is normalized using the min-max method and the processed data is recorded as x input =[V max ,V min ,V pack ], where V max 、V min 、V pack They represent the maximum voltage, minimum voltage and real-time total working voltage of the electric vehicle battery pack respectively.

[0047] (d) Set a sliding time window for the input three-dimensional data. The mathematical expression of the original network input data is as follows:

[0048]

[0049] t represents the tth timestamp, and w represents the sliding window size.

[0050] Step (2) Building a network model driven by the Mamba architecture, such as Figure 1 As shown in (right figure), it includes fully connected layers, feature layer regularization, gating units and selective SSM modules.

[0051] First, the data is regularized through the feature layer to obtain Input two fully connected layers in two ways:

[0052]

[0053]

[0054] Then, After the depth-separable convolutional layer, the SiLU activation function is input and the result is

[0055]

[0056] After SiLU activation function, we get

[0057]

[0058] After the selective SSM module, it enters the feature layer regularization to obtain and Perform multiplication to get

[0059]

[0060] final, After passing through the fully connected layer, it is added to the original input to get the final output X O :

[0061]

[0062] Optional SSM module construction: First, build the state space model (SSM) module, which captures the system state variables by representing the system using a set of first-order differential equations:

[0063]

[0064] y(t)=Ch(t)+Dx(t)

[0065] in, Represents the derivative of the state vector h(t). Matrices A, B, C, and D define the relationship between h(t), x(t), and y(t).

[0066] Then, since SSM cannot process discrete data, SSM is adapted to discrete data:

[0067]

[0068]

[0069] This discrete form of SSM can effectively establish the hidden state h k Related recursive relations. In the above formula, They are composed of parameter matrices. Usually used as a residual connection, it can be simplified in some cases. From continuous time signal to discrete time signal, a step size Δ is needed to transition, thereby converting the continuous signal x(t) into a discrete signal x k =x(kΔ). Thus, the following derivation formula is obtained:

[0070]

[0071]

[0072]

[0073] in,(·) -1 Represents the inverse of the matrix. For SSM, it has the inherent linear time invariant property, and the same matrix is ​​used for different inputs. This limits the ability to understand the differences between different sequences. Therefore, a selective SSM module is built. The selective SSM module will and Δ are set to dynamic elements that change according to the input, so that the selective SSM module can effectively adapt to different input context information. Specifically, the fully connected layer is embedded in the selective SSM module. and Δ for adjustability learning.

[0074] Step (3): Mamba architecture driven electric vehicle battery pack SOC and SOE joint estimation method. Two task layers are built to respectively realize the estimation of electric vehicle battery pack SOC and SOE. The features are obtained after the Mamba module. mamba Two task layers are built in series after the Mamba module. The task layers are composed of a fully connected layer, a softmax activation function, a flatten layer, a ReLU activation function, a dropout layer, and a fully connected layer in series.

[0075]

[0076] in Represent the first and second fully connected layers respectively, σ[·] represents the ReLU activation function, F t (·) represents the overall function of the entire task layer. The above operations are performed on each spatial dimension information to obtain the two prediction outputs of the final network, namely the estimated SOC and SOE values ​​of the electric vehicle battery pack at each moment.

[0077] For both tasks, the highest battery pack voltage-time, the lowest battery pack voltage-time, and the real-time total battery pack voltage-time in the data set are used as network inputs. The SOC and SOE results integrated in the corresponding vehicle-side battery management system are used as labels, and the estimation errors of SOC and SOE are calculated.

[0078]

[0079] Among them, SOX p and SOX r Represent the predicted value and true value of SOC and SOE respectively.

[0080] Step (4) The present invention conducts a comparative experiment on the SOC and SOE estimation effects of the battery pack on the actual operation data of the electric vehicle, wherein MAE and RMSE are used as evaluation indicators. The comparison method adopts the currently widely used CNN-LSTM method, and the test results are shown in Table 1. The battery SOC estimation effect is experimented on the public data set of the University of Maryland (US06 as the test set), and the results are shown in Table 2.

[0081] Table 1

[0082] Estimation Model MAE(%) RMSE(%) SOC Error(%) SOE Error(%) CNN-LSTM 1.277 1.424 2.6 2.7 The present invention 0.639 0.874 1.7 1.9

[0083] Table 2

[0084] Estimation Model MAE(%) RMSE(%) SOC Error(%) CNN-LSTM 0.973 1.117 3.3 The present invention 0.201 0.375 1.8

[0085] Among them, MAE and RMSE are mean absolute error and root mean square error loss functions, respectively, SOC Error is the error of online estimation of battery pack SOC, and SOE Error is the error of online estimation of battery pack SOE. It can be seen from Tables 1 and 2 that the electric vehicle battery pack SOC and SOE joint estimation method driven by the Mamba architecture proposed in the present invention can well realize the electric vehicle battery pack SOC and SOE joint estimation task, and can ensure high accuracy.

[0086] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A Mamba architecture driven battery pack SOC and SOE estimation method, characterized in that: The steps include: S1. Collect and preprocess the raw data of electric vehicle battery packs, and make the preprocessed battery data into a data set; the specific process of preprocessing the raw data of electric vehicle battery packs is as follows: First, all raw data are cleaned to remove outliers; Secondly, the Pearson correlation coefficient method is used to perform correlation analysis on the real-time running mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, battery pack real-time total operating voltage, battery pack current, battery pack SOC, and battery pack SOE of the electric vehicle battery pack, and the three groups of original data with the greatest correlation with the battery pack SOC and SOE are obtained, that is, the battery pack maximum voltage, battery pack minimum voltage, and battery pack real-time total operating voltage are selected as the original input data; Then, the original input data is normalized by the min-max method, and the processed data is recorded as x input =[V max ,V min ,V pack ], where V max 、V min 、V pack They represent the maximum voltage, minimum voltage and real-time total working voltage of the electric vehicle battery pack respectively; Finally, a sliding time window is set for the input three-dimensional data, and the mathematical expression of the input data is as follows: t represents the tth timestamp, w represents the sliding window size; S2. Construct a Mamba network, which includes a fully connected layer, a feature layer regularization, a gating unit, and a selective state space model SSM module. The Mamba network architecture implementation process is as follows: First, the input data is regularized through the feature layer to obtain Input two fully connected layers in two ways and get and Secondly, the gating unit includes a SiLU activation function and a multiplication operation, After the depth-separable convolutional layer, the SiLU activation function is input and the result is After SiLU activation function, we get Then, After the selective SSM module, it enters the feature layer regularization to obtain and Perform multiplication to get final, After the fully connected layer Add together to get the final output X O ; S3. Build two task layers, and use the output of the Mamba network to estimate the SOC and SOE of the electric vehicle battery pack.

2. The Mamba architecture driven battery pack SOC and SOE estimation method according to claim 1, characterized in that: The battery pack raw data in step S1 includes the real-time running mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, real-time total operating voltage, battery pack current and corresponding battery pack SOC and battery pack SOE data of the electric vehicle battery pack.

3. The Mamba architecture driven battery pack SOC and SOE estimation method according to claim 2, characterized in that: In step S1, the method for collecting raw data of the electric vehicle battery pack is as follows: the real-time operation, charging, and static data of the electric vehicle are uploaded to the cloud platform, and the real-time operation mileage, speed, external temperature, battery pack operating temperature, battery pack maximum voltage, battery pack minimum voltage, real-time total operating voltage of the battery pack, battery pack current, and corresponding battery pack SOC and battery pack SOE data of the electric vehicle battery pack are obtained from the cloud platform.

4. The Mamba architecture driven battery pack SOC and SOE estimation method according to claim 3, characterized in that: The specific implementation process of the selective SSM module is as follows: First, build the state space model SSM module, which captures the system state variables by using a set of first-order differential equations: y(t)=Ch(t)+Dx(t) in, Represents the derivative of the state vector h(t). Matrices A, B, C, and D define the relationship between h(t), x(t), and y(t). Second, the SSM is adapted to discrete data: in, They are composed of parameter matrices, As a residual connection, from the continuous time signal to the discrete time signal, using the step size Δ transition, the continuous signal x(t) is converted into a discrete signal x k =x(kΔ), and the following derivation formula is obtained: in,(·) -1 Represents the inverse of the matrix, and embeds the fully connected layer in the selective SSM module. and Δ for adjustability learning.

5. The method for estimating SOC and SOE of a battery pack driven by a Mamba architecture according to any one of claims 1 to 4, characterized in that: In step S3, the task layer construction method is: After the feature passes through the Mamba network, feature X is obtained mamba , two task layers are serially connected after the Mamba network; the task layers are composed of a fully connected layer, a softmax activation function, a flatten layer, a ReLU activation function, a dropout layer, and a fully connected layer in series, and two prediction outputs of the network are obtained, namely the estimated values ​​of SOC and SOE of the electric vehicle battery pack at each moment.

Images