A high-speed lithium battery energy state estimation method

Abstract

The present invention discloses a high-speed lithium battery energy state estimation method, which belongs to the field of electric vehicles. The method comprises: performing multi-time scale modeling and hierarchical feature extraction on the current, voltage, and power data of the lithium battery through a deep echo state network, and independently adjusting the hyperparameters of each layer using a parameter differentiation strategy to adapt to different dynamic characteristic inputs. Furthermore, a modal decomposition denoising module is used to remove sudden changes in output signals during the prediction process, thereby improving prediction accuracy and stability. Experimental verification shows that the method exhibits high-speed prediction performance, high accuracy, and high ease of use under a variety of real-world automotive operating conditions. It not only has the ability to quickly train and efficiently predict, but also effectively avoids the gradient vanishing problem in traditional recurrent neural networks while maintaining high accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of electric vehicles, and in particular relates to a high-speed lithium battery energy state estimation method. Background Art

[0002] In electric vehicle battery management systems, the State of Energy (SOE) is an important parameter for measuring the remaining available energy in the battery, which directly affects the endurance and energy management strategy of electric vehicles.

[0003] However, since the external measurable key data of lithium batteries have significant nonlinear and time-varying characteristics, how to efficiently and accurately estimate the SOE of batteries has become the main difficulty and research hotspot in current technological development. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a high-speed lithium battery energy state estimation method, comprising:

[0005] Acquiring input data of the lithium battery, wherein the input data includes current, voltage, and power;

[0006] Obtain the original data set based on general driving condition experiments;

[0007] Constructing a deep echo state network, and training the deep echo state network based on the original data set to obtain a lithium battery SOE estimation model;

[0008] Inputting the input data into the lithium battery SOE estimation model for calculation to obtain a SOE prediction value with output noise;

[0009] A modal decomposition and noise reduction module is constructed, and based on the modal decomposition and noise reduction module, a sudden change output signal in the SOE prediction value with output noise is removed to obtain a smooth lithium battery energy state prediction value.

[0010] Preferably, the deep echo state network includes a multi-layer stacked reservoir layer, through which the input data is modeled at multiple time scales and hierarchical feature extraction is performed;

[0011] The deep echo state network adopts a parameter differentiation strategy to independently adjust the hyperparameters of different reservoir layers, and each reservoir layer processes input data with different dynamic characteristics after adjustment.

[0012] Preferably, the step of updating the reserve pool state of the deep echo state network includes:

[0013] According to the input data and the initialized weight matrix, the state of the first layer reservoir is updated, the external input is combined with the bias, and the first layer state is obtained through nonlinear transformation;

[0014] The output of the first-layer storage pool is used as input to update the status of the second-layer and above storage pools, passing information layer by layer to form deep dynamics.

[0015] Preferably, the output weight training step of the deep echo state network includes:

[0016] Generate a joint state matrix based on the updated reserve pool state;

[0017] Remove transient data from the joint state matrix and retain steady-state data for training;

[0018] The output weights are calculated using the ridge regression method to obtain the SOE prediction value with output noise.

[0019] Preferably, the process of removing the sudden output signal in the SOE prediction value with output noise based on the noise reduction module includes:

[0020] Inputting the SOE prediction value with output noise into the modal decomposition and noise reduction module;

[0021] The predicted value is decomposed into multiple intrinsic mode functions and residual signals using successive variational mode decomposition technology.

[0022] Preferably, the denoising step of the modal decomposition denoising module includes:

[0023] Perform Pearson correlation analysis on each intrinsic mode function obtained by decomposition and calculate its correlation coefficient with the original predicted value;

[0024] According to the preset correlation threshold, the intrinsic mode functions with high correlation are screened out and the noise signals with low correlation are removed.

[0025] Preferably, the reconstruction step of the modal decomposition and noise reduction module includes: adding and reconstructing the screened high-correlation intrinsic mode functions to obtain a final smooth and accurate lithium battery energy state prediction value.

[0026] Preferably, the optimization process of the successive variational modal decomposition technique includes:

[0027] The augmented Lagrangian function is introduced to transform the modal decomposition problem into a constrained optimization problem;

[0028] The alternating direction multiplier method is used for iterative optimization to update the mode functions, center frequencies, and Lagrange multipliers.

[0029] On the other hand, the present invention further provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the method is implemented when the processor executes the computing program.

[0030] On the other hand, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the method when executed by a processor.

[0031] Compared with the prior art, the present invention has the following advantages and technical effects:

[0032] (1) First, a multi-layer stacked reservoir structure is designed to enhance the nonlinear expression capability, and a parameter differentiation strategy is used to explicitly control the dynamic characteristics of each layer. A deep echo state network is constructed for lithium-ion battery SOE estimation, enabling rapid training and testing of the model.

[0033] (2) A modal decomposition denoising module is introduced into the model, and the successive variational mode decomposition (SVMD) technique is adopted to decompose the network output into multiple intrinsic mode functions (IMFs). After performing Pearson correlation analysis on the IMFs, the threshold is used to filter and remove the noise signal, thereby weakening the impact of output noise on SOE estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0035] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] Example 1

[0039] like Figure 1 As shown, this embodiment provides a high-speed lithium battery energy state estimation method, including:

[0040] Acquiring input data of the lithium battery, wherein the input data includes current, voltage, and power;

[0041] Obtain the original data set based on general driving condition experiments;

[0042] Constructing a deep echo state network, and training the deep echo state network based on the original data set to obtain a lithium battery SOE estimation model;

[0043] Inputting the input data into the lithium battery SOE estimation model for calculation to obtain a SOE prediction value with output noise;

[0044] A modal decomposition and noise reduction module is constructed, and based on the modal decomposition and noise reduction module, a sudden output signal in the SOE prediction value with output noise is removed to obtain a smooth and accurate lithium battery energy state prediction value.

[0045] The concept of SOE was proposed by Keil Mamadou at the 216th ECS Conference in 2009. SOE is a complementary concept to the state of charge (SOC), representing the actual energy stored in a battery, that is, the percentage of energy stored in a battery relative to its rated energy. SOE is often used to more accurately represent the battery's state of energy storage, taking into account the battery's rated energy. SOE varies with the battery's charge and discharge process, and its range typically ranges from 0% (fully discharged) to 100% (fully charged). Its expression is shown in Equation (1).

[0046]

[0047] in is the battery rated energy, The remaining energy of the battery.

[0048] Furthermore, the deep echo state network includes a multi-layer stacked reservoir layer, through which the input data is modeled at multiple time scales and hierarchical features are extracted;

[0049] The deep echo state network adopts a parameter differentiation strategy to independently adjust the hyperparameters of different reservoir layers, and each reservoir layer processes input data with different dynamic characteristics after adjustment.

[0050] The flow chart of the method for estimating the SOE of lithium batteries proposed in this embodiment is as follows: Figure 1 As shown. The core of this method is the Deep Echo State Network (DeepESN) and the output denoising strategy. The core of DeepESN is a multi-layer stacked reservoir. First, the low-level reservoir maps the original data to convert signals such as current and voltage into a series of local dynamic features. Subsequently, the high-level reservoir combines information from the feature input reservoir to further capture the dynamic changes in signals such as current and voltage and complete the state update. DeepESN's random fixed weights in the reservoir and the linear output layer training mechanism enable DeepESN to not only have the ability to train quickly and make efficient predictions when performing SOE predictions, but also effectively avoid the gradient vanishing problem in traditional recurrent neural networks while maintaining high accuracy and robustness.

[0051] In DeepESN construction, the first step is to input the weight matrix and the inter-layer weight matrix The initialization of is shown in Formula 2. The role of the input weight matrix is ​​to initialize the weight of the input to the first layer of the reservoir, controlling the influence of the input signal. The role of the inter-layer weight matrix is ​​to initialize the The layer receives the weight of the previous layer's output and controls the strength of information transmission between layers.

[0052]

[0053] in Represents the number of neurons in each layer of the reserve pool, represents the input dimension, represents the input scaling factor, Inter-layer connection scaling factor, Represents the norm operation, which is used to normalize the norm of the random matrix to the unit value, ensuring that the initialization of the weight matrix has good numerical stability and dynamic behavior.

[0054] Then the weight matrix is ​​circulated The initialization of is shown in formulas (3) and (4). The loop weight matrix ensures the stability of the loop dynamics by introducing a leakage rate to balance the historical state with the current input.

[0055]

[0056]

[0057] in For the The original random matrix of the layer recurrent weight matrix, is the preset spectrum radius, is the matrix spectral radius calculation, is the scaling matrix after adjusting the spectral radius, is the leakage rate, is the identity matrix.

[0058] Furthermore, the step of updating the reserve pool state of the deep echo state network includes:

[0059] According to the input data and the initialized weight matrix, the state of the first layer reservoir is updated, the external input is combined with the bias, and the first layer state is obtained through nonlinear transformation;

[0060] The output of the first-layer storage pool is used as input to update the status of the second-layer and above storage pools, passing information layer by layer to form deep dynamics.

[0061] Updating the reservoir state is the core step in reservoir calculation. Its significance lies in leveraging the dynamic characteristics and high-dimensional feature space of the reservoir to capture the temporal dependencies, nonlinear characteristics, and dynamic patterns of the input signal. Because this embodiment uses a parameter differentiation strategy to control the dynamic characteristics of each layer, the state update formulas for each layer are slightly different. The state update formula for the first layer combines the external input with the bias and updates the first layer state through a nonlinear transformation, as shown in Formula (5).

[0062]

[0063] in Represents the state vector at time step 1, which is used to characterize the dynamic memory of the network at time step 1. For external input, is the input bias, for Activation function, which transforms linear input into nonlinear mapping, enhances the reservoir's ability to represent complex input patterns. is the circulant weight matrix, Indicates that the reservoir is in the previous time step The state vector at time .

[0064] No. The state update of a layer is achieved by using the output of the previous layer as input and passing information layer by layer to form deep dynamics, as shown in formula (6).

[0065]

[0066] in Indicates that at time step The state vector at time , Indicates in Layer at time step The output state of The input signal of the layer, Indicates in Layer at time step The state vector of .

[0067] Furthermore, the output weight training step of the deep echo state network includes:

[0068] Generate a joint state matrix based on the updated reserve pool state;

[0069] Remove transient data from the joint state matrix and retain steady-state data for training;

[0070] The output weights are calculated using the ridge regression method to obtain the SOE prediction value with output noise.

[0071] As shown in formula (7), after the state is updated, a joint state matrix will be generated , used for linear regression in the output layer, where The number of layers of deep echo state.

[0072]

[0073] Represents the state value of the deepest layer (layer N) in the reserve pool network at the last moment T.

[0074] After generating the joint state matrix, transient states need to be removed, as shown in formula (8). Removing transient states is a key step in the DeepESN training process. The purpose is to discard the unstable state during the initial operation of the network and ensure that only valid steady-state data is used in subsequent training.

[0075]

[0076] in is the joint state matrix, is the target output matrix, is the transient duration, is the total time step, is the training input matrix after removing transients, is the training target matrix after removing the transient state.

[0077] Finally, the output weight training and joint prediction of the reserve pool are performed. As shown in formula (9), the deep echo state network uses Ridge Regression to calculate the output weight ,in is the regularization coefficient, which prevents overfitting by minimizing the output error. The concatenated multi-layer state is then combined with the bias to generate the predicted output through linear transformation. , as shown in formula (10).

[0078]

[0079]

[0080] Furthermore, the process of removing the sudden output signal in the SOE prediction value with output noise based on the noise reduction module includes:

[0081] Inputting the SOE prediction value with output noise into the modal decomposition and noise reduction module;

[0082] The predicted value is decomposed into multiple intrinsic mode functions and residual signals using successive variational mode decomposition technology.

[0083] Furthermore, the denoising step of the modal decomposition denoising module includes:

[0084] Perform Pearson correlation analysis on each intrinsic mode function obtained by decomposition and calculate its correlation coefficient with the original predicted value;

[0085] According to the preset correlation threshold, the intrinsic mode functions with high correlation are screened out and the noise signals with low correlation are removed.

[0086] Furthermore, the reconstruction step of the modal decomposition and noise reduction module includes: adding and reconstructing the screened high-correlation intrinsic mode functions to obtain a final smoothed lithium battery energy state prediction value.

[0087] Furthermore, the optimization process of the successive variational modal decomposition technique includes:

[0088] The augmented Lagrangian function is introduced to transform the modal decomposition problem into a constrained optimization problem;

[0089] The alternating direction multiplier method is used for iterative optimization to update the mode functions, center frequencies, and Lagrange multipliers.

[0090] After obtaining the SOE prediction value with output noise generated by the deep recurrent state network, a denoising module is designed to remove the output noise in order to obtain a smooth and stable prediction result. The core of the denoising module is the successive variational mode decomposition (SVMD). As shown in formulas (11) and (12), SVMD can transform the original signal into Decompose into current mode and the residual signal , where the residual signal contains the extracted mode set and the untreated residuals .

[0091]

[0092]

[0093] The goal of SVMD is to minimize the combination of three criteria while satisfying the signal constraints, as shown in equations (13) to (15).

[0094]

[0095]

[0096]

[0097] in is the spectrum compactness, is the separation of residual signal and modal spectrum, is the spectrum separation between modes. In formula (13) is the time derivative operator, is the unit impulse function, is the time domain signal, is the center frequency of the current mode. In formula (14), is the filter of the residual signal, which realizes spectrum filtering through the time domain equivalent operation of the frequency domain product. In formula (15), It is an inter-modal separation filter that forces the spectrum of the current mode to be away from the center frequencies of all extracted modes through multiple groups of filters to avoid modal aliasing.

[0098] In order to solve this optimization problem, the augmented Lagrangian function is introduced. The augmented Lagrangian function transforms the original optimization problem into a constrained optimization problem, and ensures accurate signal reconstruction through the penalty term and Lagrangian multiplier, as shown in formula (16).

[0099]

[0100] in is the regularization parameter, is the Lagrange multiplier, which ensures the integrity of signal decomposition.

[0101] SVMD uses the alternating direction multiplier method (ADMM) optimization to iteratively update the following variables. First, each mode is updated in the frequency domain. Each mode is updated in the form of a frequency domain filter, and the influence of other modes is excluded during calculation, as shown in formula (17).

[0102]

[0103] in is the frequency domain update value of the current modal component, is the frequency domain representation of the original signal, is the current center frequency, is the current iteration mode, is the frequency domain representation of the Lagrange multiplier.

[0104] Next, update the center frequency of each mode as shown in formula (18), where is the updated center frequency.

[0105]

[0106] Then, the dual ascent method is used to update the Lagrange multiplier to ensure that the signal reconstruction constraints are gradually satisfied during the optimization process, as shown in formula (19).

[0107]

[0108] in is the dual ascent step size, which is used to control the convergence speed.

[0109] Finally, the convergence is determined by checking the modal convergence and residual energy conditions, as shown in formulas (20) and (21). If it has not converged, the alternating direction multiplier method is continued for iterative update. Otherwise, the decomposition result is output.

[0110]

[0111]

[0112] in is the modal convergence threshold, which is used to control the relative error of modal update. is the residual convergence threshold, which controls the deviation of the residual energy from the preset threshold. represents the estimated value of the Lth mode after the n+1th iteration, represents the estimated value of the Lth mode after the nth iteration, is the preset residual energy threshold.

[0113] When SVMD reaches convergence, it outputs multiple intrinsic mode functions (IMFs), and then calculates the Pearson correlation coefficient according to formula (22) to measure the degree of linear correlation between the DeepESN network prediction results and the IMFs.

[0114]

[0115] in, is the Pearson correlation coefficient, is the decomposed signal IMF, is the prediction result of DeepESN neural network, is the sample size, is the IMF mean of the decomposed signal, is the mean of the predicted results.

[0116] After obtaining the correlation coefficient between each IMF and the forecast result, we set the threshold The correlation between each modal component and the original signal is determined to determine the noise content. The goal is to ensure that components with high correlation are considered valid signals, while components with low correlation are considered noise. In this embodiment, the correlation threshold is set to 0.6.

[0117] After screening out the high correlation intrinsic mode function (HIMF), the signal is added and reconstructed, and finally a smoother and more accurate SOE prediction value is obtained, as shown in formula (23).

[0118]

[0119] in is the number of IMFs greater than the threshold, Indicates the A highly correlated intrinsic mode function.

[0120] On the other hand, this embodiment further provides an electronic device, comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the method is implemented when the processor executes the computing program.

[0121] On the other hand, this embodiment further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the method when executed by a processor.

[0122] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A high-speed lithium battery energy state estimation method, characterized in that: include: Acquiring input data of the lithium battery, wherein the input data includes current, voltage, and power; Obtain the original data set based on general driving condition experiments; Constructing a deep echo state network, and training the deep echo state network based on the original data set to obtain a lithium battery SOE estimation model; Inputting the input data into the lithium battery SOE estimation model for calculation to obtain a SOE prediction value with output noise; Constructing a modal decomposition noise reduction module, and removing the sudden output signal in the SOE prediction value with output noise based on the modal decomposition noise reduction module to obtain a smooth lithium battery energy state prediction value; The deep echo state network includes a multi-layer stacked reservoir layer, through which the input data is modeled at multiple time scales and hierarchical features are extracted; The deep echo state network adopts a parameter differentiation strategy to independently adjust the hyperparameters of different reservoir layers, and each reservoir layer after adjustment processes input data with different dynamic characteristics.

2. The method according to claim 1, characterized in that The step of updating the reserve pool state of the deep echo state network includes: According to the input data and the initialized weight matrix, the state of the first layer reservoir is updated, the external input is combined with the bias, and the first layer state is obtained through nonlinear transformation; The output of the first-layer storage pool is used as input to update the status of the second-layer and above storage pools, passing information layer by layer to form deep dynamics.

3. The method according to claim 2, characterized in that The output weight training step of the deep echo state network includes: Generate a joint state matrix based on the updated reserve pool state; Remove transient data from the joint state matrix and retain steady-state data for training; The output weights are calculated using the ridge regression method to obtain the SOE prediction value with output noise.

4. The method according to claim 3, characterized in that The process of removing the sudden output signal in the SOE prediction value with output noise based on the noise reduction module includes: Inputting the SOE prediction value with output noise into the modal decomposition and noise reduction module; The SOE prediction value is decomposed into multiple intrinsic mode functions and residual signals using successive variational mode decomposition technology.

5. The method according to claim 1, wherein The denoising steps of the modal decomposition denoising module include: Perform Pearson correlation analysis on each intrinsic mode function obtained by decomposition and calculate its correlation coefficient with the original predicted value; According to the preset correlation threshold, the intrinsic mode functions with high correlation are screened out and the noise signals with low correlation are removed.

6. The method according to claim 1, characterized in that The reconstruction step of the modal decomposition and noise reduction module includes: adding and reconstructing the screened high-correlation intrinsic mode functions to obtain a final smooth and accurate lithium battery energy state prediction value.

7. The method according to claim 4, characterized in that The optimization process of the successive variational modal decomposition technique includes: The augmented Lagrangian function is introduced to transform the modal decomposition problem into a constrained optimization problem; The alternating direction multiplier method is used for iterative optimization to update the mode functions, center frequencies, and Lagrange multipliers.

8. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein: When the processor executes the computing program, the method according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images