SOC estimation method commonly used for sodium battery and lithium battery

Abstract

The invention discloses an SOC estimation method commonly used for a sodium battery and a lithium battery. The SOC estimation method comprises the steps of obtaining sodium battery charging and discharging characteristic data and lithium battery charging and discharging characteristic data; the sodium battery charging and discharging characteristic data and the lithium battery charging and discharging characteristic data are sent to an LN-HNet model; wherein a network layer of the model performs deep feature extraction and gradient optimization on sodium battery conversion data and lithium battery conversion data through two residual modules, modeling is performed in combination with heterogeneous features of an HNet module, and finally multi-source data are spliced and fused through channels; the channel connection layer is used for splicing the processed feature data from different paths together according to channel dimensions; the spliced feature data passes through a root-mean-square normalization layer and a Sigmoid activation function in sequence, and an SOC estimated value is output; the model finds a common rule of the two battery types in the SOC change process through comprehensive analysis of the characteristics, and improves the accuracy through learning specific response modes of different batteries.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and in particular to a SOC estimation method commonly used for sodium batteries and lithium batteries. Background Art

[0002] Battery state of charge (SOC) is a key indicator for measuring the remaining power of the battery, which directly affects the health management and performance optimization of the battery. In particular, in electric vehicles and renewable energy storage systems, accurate SOC estimation is crucial to ensuring the reliability and economy of the battery. At present, the industry has developed a variety of solutions for SOC estimation methods, mainly including open circuit voltage method (OCV), ampere-hour integration method, model-based filtering method and data-driven method.

[0003] The open circuit voltage method (OCV) is based on the relationship between the battery open circuit voltage and SOC, and can provide a relatively accurate estimate, but it is only applicable to static conditions and requires the battery to be stationary for a long time to eliminate the load effect. The ampere-hour integration method estimates the SOC by calculating the battery's charge and discharge current in real time. Its advantage is strong real-time performance, but it is easily affected by cumulative errors, especially when the battery ages or the temperature changes, the accuracy decreases. Model-based filtering methods, such as Kalman filtering, combine the battery's mathematical model with real-time data for estimation, and can cope with dynamic changes, but accurate calibration and model complexity are its challenges. Data-driven methods use machine learning technology to predict SOC by learning historical data. Although it can improve the estimation accuracy, its performance is limited by the quality of training data and lacks physical explanation.

[0004] In response to the SOC estimation needs of different types of batteries (such as lithium batteries and sodium batteries), existing technologies often require targeted design, which increases the complexity and cost of the BMS system. Lithium batteries often use complex electrochemical models and open circuit voltage measurements, while sodium batteries need to consider their unique reaction mechanisms and kinetic responses. In addition, most of the current models solve the SOC estimation of a single type of battery, and do not consider batteries made of different materials. Summary of the invention

[0005] In view of the problems in the prior art, the present invention provides a SOC estimation method that is commonly applicable to sodium batteries and lithium batteries, with the purpose of being able to perform SOC estimation across the boundaries of battery types.

[0006] A SOC estimation method commonly used for sodium batteries and lithium batteries comprises the following steps:

[0007] Step 1: Obtaining sodium battery charge and discharge characteristic data and lithium battery charge and discharge characteristic data;

[0008] Step 2: Feed the charge and discharge characteristic data of the sodium battery and the charge and discharge characteristic data of the lithium battery into the LN-HNet model, which includes two 1×1 convolutional layers, a network layer, a channel connection layer, a root mean square normalization layer, and a Sigmoid activation function;

[0009] Step 3: The 1×1 convolutional layer is used to perform channel-level transformation on the input data and obtain the transformed data of the sodium battery and the transformed data of the lithium battery;

[0010] Step 4: The network layer performs deep feature extraction and gradient optimization on the transformed data of the sodium battery and the transformed data of the lithium battery through two residual modules respectively, and combines the heterogeneous feature modeling of the HNet module. Finally, multi-source data is fused through channel splicing;

[0011] Step 6: The channel connection layer concatenates the processed feature data from different paths together along the channel dimension;

[0012] Step 7: The concatenated feature data passes through the root mean square normalization layer and the Sigmoid activation function in sequence, and the SOC estimation value is output.

[0013] Furthermore: The heterogeneous feature modeling includes channel / space aggregation, multi-scale Inception branches, and bidirectional SSM state transfer.

[0014] Furthermore: The heterogeneous feature modeling includes two main paths, and the structures of these two main paths are the same and both include a first branch, a second branch, and a third branch; the outputs of the two residual modules are respectively fed into these two main paths;

[0015] The output of the residual module enters the skip connection within the first branch;

[0016] The output of the residual module passes through channel aggregation, space aggregation, and SiLU activation function in sequence within the second branch, and deep features of the battery in different dimensions are obtained;

[0017] The output of the residual module passes through a linear transformation within the third branch and then through 1×1 convolution, 3×3 convolution, 5×5 convolution, and max pooling for feature fusion, obtaining a joint representation containing multi-scale feature information. The joint representation is respectively fed into the branch and the branch to obtain the forward SSM data and the backward SSM data respectively. The deep features are respectively multiplied with the forward SSM data and the backward SSM data through matrix multiplication. The data after the operation is then added element by element. The added and combined data passes through a linear layer and an LN layer in sequence and then is added element by element with the data of the skip connection within the first branch to obtain the output of this main path; among them, the two main paths share and transfer the state transfer matrix features of each other.

[0018] Furthermore, it is: Branch: Update the hidden state at the current moment using the forward state transition matrix. The formula is as shown: ; ;

[0019] Among them, is the hidden state at the current moment, is the forward state transition matrix, is the input matrix, is the output matrix, is the input at the current moment, is the output at the current moment:

[0020] Branch:

[0021] Update the hidden state using the backward state transition matrix. The formula is as shown: ; ;

[0022] Among them, is the backward hidden state at the current moment, is the backward state transition matrix, is the input matrix, is the output matrix, is the input at the current moment, is the output at the current moment.

[0023] Furthermore, it is: The two main paths share and transfer the characteristics of their respective state transition matrices with each other. Specifically:

[0024] Collect the data generated during the processing of these two main paths and collectively call them the sodium battery data stream and the lithium battery data stream respectively. The forward state update processes of the two main paths are respectively expressed as: ; ;

[0025] Among them, is the forward state transition matrix, and They are the input matrix and the output matrix in the main path where the sodium battery data stream is located, respectively; and They are the input matrix and the output matrix in the main path where the lithium battery data stream is located, respectively;

[0026] The backward state update processes of the two main paths are respectively expressed as: (12); (13);

[0027] Among them, is the backward state transition matrix, and They are the input matrix and the output matrix in the main path where the sodium battery data stream is located, respectively; and They are the input matrix and the output matrix in the main path where the lithium battery data stream is located, respectively.

[0028] Furthermore, the residual module includes a convolutional layer, batch normalization, an activation function, and a Dropout layer. The input data is locally processed through the convolutional layer to obtain local temporal features. One path of the local temporal features sequentially passes through batch normalization, the ReLU activation function, and the Dropout layer and then makes a skip connection with its other path to obtain the output of the residual module.

[0029] Furthermore, Step 1 includes the following steps:

[0030] Step 1.1: In an incubator at a constant temperature of 25°C, test the lithium battery and the sodium battery under three dynamic driving cycles, namely NEDC, WLTC, and CLTC-P, to obtain the original data of the sodium battery's charge and discharge and the original data of the lithium battery's charge and discharge;

[0031] Step 1.2: Perform normalization preprocessing on the original data of the sodium battery's charge and discharge and the original data of the lithium battery's charge and discharge to unify the dimensions of the characteristic parameters;

[0032] Step 1.3: Obtain the characteristic data of the sodium battery's charge and discharge and the characteristic data of the lithium battery's charge and discharge.

[0033] Furthermore, the characteristic parameters in the original data of the sodium battery's charge and discharge and the original data of the lithium battery's charge and discharge include current, voltage, and temperature.

[0034] Advantages of the present invention: By fusing the battery time series features through a convolutional neural network, it helps the model better understand the behavior of the battery. Since the deep learning model does not rely on the physical model of the battery but uses a data-driven method to accurately estimate the SOC, the LN-HNet model can extract common features from the characteristic data of lithium batteries and sodium batteries. Through comprehensive analysis of these features, the model finds the common laws of the two battery types during the SOC change process and improves the accuracy by learning the specific response patterns of different batteries. Brief Description of the Drawings

[0035] Figure 1 is a flowchart of the present invention;

[0036] Figure 2 is a structural diagram of the LN-HNet model;

[0037] Figure 3 is a structural diagram of the residual module;

[0038] Figure 4 is a structural diagram of the HNet module;

[0039] Figure 5 is a structural diagram of the Inception Block module;

[0040] Figure 6 is a graph of the estimation results of 18650Li battery;

[0041] Figure 7 is a graph of the error results of 18650Li battery;

[0042] Figure 8 is a graph of the estimation results of 26700Na battery;

[0043] Figure 9 is a graph of the error results of 26700Na battery;

[0044] Figure 10 is a graph of the estimation results of 18650Na battery;

[0045] Figure 11 is a graph of the error results of 18650Na battery. Detailed Embodiments

[0046] The present invention will be described in detail below with reference to the accompanying drawings. Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described by referring to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention. The orientation terms such as left, middle, right, up, and down in the embodiments of the present invention are only relative concepts to each other or are referenced based on the normal use state of the product, and should not be considered restrictive.

[0047] A method for estimating the state of charge (SOC) applicable to both sodium batteries and lithium batteries, comprising the following steps:

[0048] Step 1: In a 25°C constant temperature chamber, obtain the charge and discharge characteristic data of sodium batteries and the charge and discharge characteristic data of lithium batteries;

[0049] Step 1.1: Select three cylindrical batteries for standby testing; namely: a lithium battery of model 18650, a sodium battery of model 18650, and a sodium battery of model 26700. The detailed parameters of the three batteries are shown in Table 1; Table 1 Battery detailed parameters

[0050] Set the test steps of the three batteries. The specific detailed steps are shown in Table 2, Table 3, and Table 4. Among them, NEDC represents the New European Driving Cycle, WLTC represents the Worldwide Harmonized Light Vehicles Test Cycle, and CLTC-P represents the China Passenger Car Driving Cycle. These three dynamic working conditions are used to fully demonstrate the operation state of the power battery in the real environment; obtain the original charge and discharge data of sodium batteries and the original charge and discharge data of lithium batteries; Table 2 Test steps for 18650 Li battery

[0051] Table 3 Test steps for 18650 Na battery

[0052] Table 4 Test steps for 26700 Na battery

[0053] Step 1.2: Perform normalization preprocessing on the original data of sodium battery charge and discharge and the original data of lithium battery charge and discharge to unify the dimensions between characteristic parameters; in the estimation of power battery SOC, data preprocessing of the original data is to improve the estimation accuracy and the stability of the algorithm; the present invention uses Min-Max Normalization to preprocess the three measured original data, and the characteristic parameters in the original data of sodium battery charge and discharge and the original data of lithium battery charge and discharge include current, voltage and temperature; in the data-driven deep learning method, the most commonly used characteristic parameters for battery SOC estimation are current, voltage and temperature; Min-Max Normalization linearly maps the original data to a specified interval [-1,1], and its main purpose is to compress the data into the same dimension range, eliminate the differences between dimensions, and is suitable for the case where the input data varies greatly, which helps to improve the convergence speed and stability of the model; the formula is as follows: ;

[0054] wherein, is the original data value, is the minimum value in the dataset, is the maximum value in the dataset, is the normalized data value, and are respectively the upper and lower limits [-1, 1] of the normalized data range;

[0055] Step 1.3: Obtain the charge and discharge characteristic data of sodium battery and the charge and discharge characteristic data of lithium battery;

[0056] Step 2: Feed the charge and discharge characteristic data of sodium battery and the charge and discharge characteristic data of lithium battery into the LN-HNet model. The LN-HNet model includes 2 1×1 convolutional layers, a network layer, a channel connection layer, a root mean square normalization layer and a Sigmoid activation function;

[0057] Step 3: The 1×1 convolutional layer is used to perform channel-level conversion on the input data and obtain the converted data of sodium battery and the converted data of lithium battery; the 1×1 convolutional layer is a special convolutional layer, where the size of the convolutional kernel is 1×1, which enables this layer to perform channel-level conversion on the input data without changing the input spatial dimensions; during the battery SOC estimation process, the interaction of different characteristics (such as voltage, current and temperature) affects the accurate estimation of SOC; the 1×1 convolution helps the model better understand the behavior of the battery by fusing these characteristics;

[0058] Step 4: The network layer performs deep feature extraction and gradient optimization on the sodium battery conversion data and lithium battery conversion data respectively through two residual modules, and combines the heterogeneous feature modeling of the HNet module. Finally, multi-source data is fused through channel splicing.

[0059] The residual module includes a convolutional layer, batch normalization, an activation function, and a Dropout layer. The convolutional layer locally processes the input data to obtain local temporal features. In battery SOC estimation, the convolutional layer can capture the change patterns of the battery under different working conditions. The local temporal features pass through the batch normalization layer, ReLU activation function, and Dropout layer in sequence and then make a skip connection with the other path to obtain the output of the residual module. Among them, the batch normalization layer standardizes the output of the convolutional layer, reducing the internal covariance shift and making the input data of each layer more stable. This helps to accelerate the training of the model and improve the stability during the training process. The residual module also introduces skip connections, allowing the input signal to bypass some neural network layers and directly pass to the subsequent layers, avoiding the common gradient vanishing problem in deep networks. ReLU is a commonly used non-linear activation function that sets negative values to zero and keeps positive values unchanged, enabling the model to learn non-linear features. The formula is as shown; ReLU helps the model capture more complex patterns, especially in the dynamic changes of battery SOC. The Dropout layer randomly discards some neurons to avoid overfitting of the model during training. For battery SOC estimation, Dropout can help the model avoid over-reliance on certain features when processing diverse working condition data, thereby improving the generalization ability of the model. ;

[0060] Among them, is the input value;

[0061] Heterogeneous feature modeling includes channel / space aggregation, multi-scale Inception branches, and bidirectional SSM state transition. Heterogeneous feature modeling includes two main paths, and the structures of these two main paths are the same and both include a first branch, a second branch, and a third branch. The outputs of the two residual modules are respectively fed into these two main paths.

[0062] The output of the residual module enters the skip connection within the first branch.

[0063] The output of the residual module is successively passed through channel aggregation, spatial aggregation, and the SiLU activation function within the second branch to obtain the deep features of the battery in different dimensions. Channel and Spatial aggregations are regarded as part of the underlying feature extraction, and the SiLU activation function is used to enhance the non-linear representation ability, which is beneficial to extracting the deep features of the battery in different dimensions.

[0064] Among them, channel aggregation mainly adjusts the importance of each channel dynamically by weighting the channel dimension of the feature data. Each channel represents a specific feature dimension of the input feature data. By weighting these channels, channel aggregation enhances the attention to useful features and suppresses unimportant features. Its advantage lies in helping the network to adaptively select information according to the features of the input data. Channel aggregation helps to focus on the channel features that are most critical for SOC estimation and improve the accuracy of the model. The specific operation is as follows: First, the input feature data undergoes global average pooling (AvgPool), which aggregates the spatial information of each channel into a single value to obtain the global feature of each channel. Then, the feature data passes through a 1×1 convolutional layer (1x1 Conv), which compresses the number of channels of the feature data and allows the model to learn the relationships between channels. The feature data after the 1x1 convolution passes through the ReLU activation function, which sets negative values to 0 and retains positive values to help the model learn non-linear features. Then, the data passes through another 1×1 convolutional layer for another feature transformation, mapping the feature data to new spatial features. Spatial aggregation enhances the attention to key information regions by weighting the spatial positions of the feature data. Different from channel aggregation, spatial aggregation focuses on the spatial dimension of the feature data and automatically pays attention to which regions in the input data are most important for SOC estimation. The specific operation is as follows: The input feature data undergoes max pooling (MaxPool), which extracts the maximum value in the feature data to highlight the important regional features in the feature data. Then, the feature data passes through a 7×7 convolutional layer, which captures the spatial up-and-down relationships of the input data through a relatively large convolutional kernel and extracts the local dependence information of the space to help the network understand the spatial features. The feature data after convolution enters the Sigmoid activation function, which compresses the attention of the spatial position between 0 and 1 to further reflect the importance of each spatial position in the input feature data. Finally, the obtained spatial attention features are multiplied by the original input features to weight each spatial position, thereby enhancing the features of important spatial regions and suppressing the features of unimportant regions. The data after feature extraction passes through the SiLU activation function, which is a smooth activation function that can further enhance non-linear features without losing gradient information. The formula is as shown; ;

[0065] Among them, is the Sigmoid function, and the formula is: ;

[0066] Among them, is the input value;

[0067] The output of the residual module is linearly transformed in the third branch and then undergoes feature fusion after 1×1 convolution, 3×3 convolution, 5×5 convolution, and max pooling in the Inception Block module. These four parallel sub-branches are respectively used to extract feature information under different receptive fields to obtain a joint representation containing multi-scale feature information. Among them, the 1×1 convolution operation is used for linear projection and channel compression, which can reduce the computational complexity while retaining key information; the 3×3 convolution operation is used to extract local spatio-temporal features under medium receptive fields to enhance the model's sensitivity to local SOC dynamic changes; the 5×5 convolution operation is used to model features within a larger receptive field range to capture the SOC change trend on a longer time scale; the max pooling operation is used to retain the most significant features in each local window, improve the robustness of feature selection, and have a certain noise reduction ability.

[0068] The joint representation is respectively sent into branch and branch to obtain forward SSM data and backward SSM data respectively. The deep features are respectively subjected to matrix multiplication operations with the forward SSM data and the backward SSM data. The data after the operation are then added element by element. The combined data after addition are successively passed through a linear layer and an LN layer and then added element by element with the data of the skip connection in the first branch to obtain the output of the main path, thereby enhancing the model's perception ability of the SOC state change of the sodium battery under complex working conditions. Specifically: and are respectively composed of the forward matrix , , and the backward matrix , , ; The state transition matrices and have the greatest impact on the system because they control the evolution of the current hidden state, while , , and , mainly affect the input and output states;

[0069] Branch:

[0070] Update the hidden state at the current time using the forward state transition matrix ( ), and the formula is as shown: ; ;

[0071] Among them, is the hidden state at the current time, is the forward state transition matrix, is the input matrix, is the output matrix, is the input at the current time, is the output at the current time.

[0072] Branch:

[0073] Update the hidden state using the backward state transition matrix ( ), and the formula is as shown: ; ;

[0074] Among them, is the backward hidden state at the current time, is the backward state transition matrix, is the input matrix, is the output matrix, is the input at the current time, is the output at the current time;

[0075] Among them, the two main paths share and transfer the characteristics of their respective state transition matrices with each other. That is, the data generated during the processing of these two main paths are collectively referred to as the sodium battery data stream and the lithium battery data stream respectively. The forward state update processes of the two main paths are respectively expressed as: ; ;

[0076] Among them, is the forward state transition matrix, and are the input matrix and the output matrix in the main path where the sodium battery data stream is located respectively; and are the input matrix and output matrix in the main path where the lithium battery data stream is located, respectively;

[0077] The backward state update processes of the two main paths are respectively expressed as: (12); (13);

[0078] Among them, is the backward state transition matrix, and are the input matrix and output matrix in the main path where the sodium battery data stream is located, respectively; and are the input matrix and output matrix in the main path where the lithium battery data stream is located, respectively;

[0079] In addition, the LN layer normalizes the data to ensure more stable training of the network and accelerate convergence. The formula is shown in (14): (14);

[0080] Among them, is the output of the LN layer, is the input data, is the mean of the input data, is the standard deviation of the input data, is the learnable scaling parameter, is the learnable offset parameter;

[0081] Step 6: The channel connection layer concatenates the processed feature data from different paths together along the channel dimension; this process helps to fuse features from different data sources, thereby enhancing the model's adaptability to diverse inputs; the advantages of channel connection are:

[0082] Multi-source information fusion: Through Channel Concat, data from different data sources (such as data of different types of batteries) can be effectively combined together to provide a more comprehensive feature representation;

[0083] Enhance the model's expressive ability: This concatenation operation enables the network to learn more types of features and improves the model's processing ability for diverse inputs;

[0084] Improve information flow: The concatenated feature data contains information from different paths, thereby enhancing the network's perception ability, especially when dealing with complex and diverse data;

[0085] Step 7: The concatenated feature data passes through a root mean square normalization layer and a Sigmoid activation function in sequence, and the SOC estimation value is output. Among them,

[0086] The root mean square normalization layer (RMS Norm) normalizes the data, making the data more stable during the training process. The normalization formula of RMS Norm is shown in (14); (15);

[0087] Among them, is the output of the root mean square normalization layer, is the input data, is the feature dimension;

[0088] The Sigmoid function restricts the output between 0 and 1. The smoothness of the Sigmoid function makes the SOC estimation result more stable, avoiding violent fluctuations, and is suitable for representing the SOC of the battery.

[0089] These steps together ensure that the LN-HNet model can accurately and efficiently estimate the SOC of lithium-ion batteries and sodium batteries, adapting to the changes in battery characteristics under different working conditions.

[0090] Among them, the LN-HNet model needs to be trained and debugged to obtain the best hyperparameters. The specific parameters are shown in Table 5;

[0091] Table 5 Model Hyperparameter Settings

[0092] The present invention uses two evaluation metrics, RMSE and MAE, which are commonly used in regression tasks, to evaluate the performance of the model, measures the generalization ability and estimation accuracy of the model through its performance on the test set, and tests the applicability of the model on lithium-ion batteries and sodium batteries. In order to verify the effectiveness of the LN-HNet model in estimating the SOC of batteries with different material types, the present invention uses the remaining one working condition data of three batteries for verification, and the ambient temperature is 25°C; Figures 6 to 11 are the estimation and error results of the three batteries under different working conditions; Table 6 is the evaluation results of the LN-HNet model on the three batteries; Table 6 Evaluation Results of the LN-HNet Model on Three Batteries

[0093] From the evaluation results in the table, it can be seen that the LN-HNet model shows high SOC estimation accuracy under three battery types and different working conditions; the errors of the model on lithium batteries and sodium batteries are very small, especially in the two indicators of RMSE and MAE, indicating that LN-HNet can effectively adapt to different types of batteries and test conditions; generally speaking, the LN-HNet model can accurately estimate the SOC under different battery conditions and can be applied to both lithium batteries and sodium batteries at the same time, which provides a very effective solution for the battery management system.

[0094] To verify the generalization and applicability of the proposed model, a mainstream time series estimation model was used for comparative model testing; Table 7 shows that the RMSE and MAE values of the proposed model under all test conditions and battery models are lower than those of Mamba and TCN-LSTM, indicating that there are obvious performance differences among the three models, and the proposed model shows higher estimation accuracy.

[0095] Table 7 Comparative evaluation results of three models

[0096] In summary, the LN-HNet model performs optimally in all evaluation indicators. It not only far exceeds Mamba and TCN-LSTM in terms of SOC estimation accuracy, but also demonstrates good performance in model generalization and robustness, verifying the effectiveness and wide applicability of its network structure design and optimization strategy in practical applications.

[0097] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A SOC estimation method commonly used for sodium batteries and lithium batteries, characterized in that: The following steps are involved: Step 1: Obtaining sodium battery charge and discharge characteristic data and lithium battery charge and discharge characteristic data; Step 2: The sodium battery charge and discharge characteristic data and the lithium battery charge and discharge characteristic data are sent to the LN-HNet model. The LN-HNet model includes two 1×1 convolutional layers, a network layer, a channel connection layer, a root mean square normalization layer, and a Sigmoid activation function; Step 3: Use a 1×1 convolutional layer to convert the input data at the channel level and obtain sodium battery conversion data and lithium battery conversion data; Step 4: The network layer uses two residual modules to perform deep feature extraction and gradient optimization on the sodium battery conversion data and the lithium battery conversion data respectively, and combines the heterogeneous feature modeling of the HNet module, and finally merges the multi-source data through channel splicing; Step 6: The channel connection layer concatenates the processed feature data from different paths according to the channel dimension; Step 7: The concatenated feature data is passed through the RMS normalization layer and the Sigmoid activation function in turn, and the SOC estimation value is output.

2. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 1, characterized in that: Heterogeneous feature modeling includes channel / spatial aggregation, multi-scale Inception branches, and bidirectional SSM state transfer.

3. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 2, characterized in that: The heterogeneous feature modeling includes two main paths, which have the same structure and both include a first branch, a second branch and a third branch; the outputs of the two residual modules are respectively sent to the two main paths; The output of the residual module enters the skip connection in the first branch; The output of the residual module is sequentially subjected to channel aggregation, spatial aggregation, and SiLU activation function in the second branch to obtain the deep features of the battery in different dimensions; The output of the residual module is linearly transformed in the third branch and then subjected to 1×1 convolution, 3×3 convolution, 5×5 convolution and maximum pooling for feature fusion to obtain a joint representation containing multi-scale feature information, which is then sent to Branch and The forward SSM data and the backward SSM data are obtained respectively, and the deep features are respectively subjected to matrix multiplication operation with the forward SSM data and the backward SSM data. The data after the operation are then added element by element. The added and combined data are successively passed through the linear layer and the LN layer, and then added element by element with the data of the jump connection in the first branch to obtain the output of the main path; wherein, the two main paths share and transmit their respective state transfer matrix features with each other.

4. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 3, characterized in that: Branch: Use the forward state transition matrix to update the hidden state at the current moment. The formula is as follows As shown: ； ； in, is the hidden state at the current moment, is the forward state transition matrix, is the input matrix, is the output matrix, is the input at the current moment, is the output at the current moment, Branch: Use the backward state transition matrix to update the hidden state, the formula is as follows As shown: ； ； in, is the backward hidden state at the current moment, is the backward state transition matrix, is the input matrix, is the output matrix, is the input at the current moment, is the output at the current moment.

5. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 4, characterized in that: The two main roads share and transfer their respective state transfer matrix characteristics: The data generated during the processing of these two main paths are collectively referred to as sodium battery data flow and lithium battery data flow, and the forward state update processes of the two main paths are respectively expressed as: ； ； in, is the forward state transfer matrix, and They are the input matrix and output matrix in the main path where the sodium battery data flow is located; and They are respectively the input matrix and the output matrix in the main path where the lithium battery data flow is located; The backward state update process of the two main roads is expressed as follows: （12）； （13）； in, is the backward state transfer matrix, and They are the input matrix and output matrix in the main path where the sodium battery data flow is located; and They are respectively the input matrix and output matrix in the main path where the lithium battery data flow is located.

6. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 1, characterized in that: The residual module includes a convolutional layer, batch normalization, an activation function, and a Dropout layer. The convolutional layer is used to locally process the input data and obtain local time series features. The local time series features are sequentially subjected to batch normalization, ReLU activation function, and Dropout layer, and then jump-connected with the other path to obtain the output of the residual module.

7. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: In a 25°C constant temperature box, the lithium battery and the sodium battery are tested under three dynamic conditions: NEDC, WLTC and CLTC-P, to obtain the raw data of the sodium battery charge and discharge and the raw data of the lithium battery charge and discharge; Step 1.2: Perform normalization preprocessing on the raw data of sodium battery charging and discharging and the raw data of lithium battery charging and discharging to unify the dimensions of the characteristic parameters; Step 1.3: Obtain sodium battery charge and discharge characteristic data and lithium battery charge and discharge characteristic data.

8. The SOC estimation method commonly used for sodium batteries and lithium batteries according to claim 7, characterized in that: The characteristic parameters in the raw data of sodium battery charging and discharging and the raw data of lithium battery charging and discharging include current, voltage and temperature.

Images