Battery SOH (state of health) and SOC (state of charge) combined prediction method based on neural network

Abstract

The invention relates to the field of battery SOH (state of health) prediction, in particular to a neural network-based battery SOH and SOC (state of charge) joint prediction method, which comprises the following steps of: combining an alternating-current impedance spectrum of a battery, battery charging and discharging data and an LSTM (Long Short Term Memory) network, and importing the data into Z-view software for fitting to obtain parameter values of a second-order RC equivalent circuit; parameter values of the second-order RC equivalent circuit and battery charging and discharging data serve as input of the LSTM network, and errors which cannot be solved due to the fact that only a neural network is used are avoided. According to the method, high-correlation input data is screened by using a Spearman rank correlation coefficient, the LSTM network training time and difficulty are reduced, and local features are extracted by using a multi-scale convolution kernel; a self-adaptive Dropout layer is introduced, and the discard rate is set in a standard mode. According to the method, the electrochemical impedance spectrum is measured by the electrochemical analyzer in the charging and discharging process, and the battery mechanism is fully fused by utilizing diversified input, so that the prediction result is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of battery state of health prediction, and particularly to a method for jointly predicting battery SOH and SOC based on a neural network. Background Art

[0002] The background of battery health management mainly comes from the applications of batteries in various fields, such as mobile devices, electric vehicles, smart homes, etc. With the expansion of battery application scenarios and the increase in demand, the requirements for battery performance and state of health are also getting higher and higher. A battery is an energy conversion device that converts chemical energy into electrical energy, and its performance and lifespan directly affect the working efficiency and stability of the device or system. In actual use, the battery will gradually age due to factors such as charge and discharge cycles, temperature changes, load changes, etc., resulting in problems such as a decrease in battery capacity, an increase in internal resistance, and an increase in safety risks. These problems will not only reduce the reliability and lifespan of the battery, but may also have a significant impact on the performance and safety of the device or system.

[0003] Battery health management is an important topic. By monitoring, analyzing, and evaluating the battery in real time, abnormal situations of the battery can be detected in time and corresponding measures can be taken to ensure the normal operation of the battery and extend its lifespan. The estimation of battery state of health generally can adopt the physical model method, the statistical model method, and the data-driven method. The first two methods require a large number of parameters, a complex modeling process, and a large amount of calculation. However, the data-driven method regards the battery parameters as a black box and trains using a large amount of battery data through machine learning. By learning the characteristics and patterns of the battery data, the prediction of the battery state of health is realized. SOH and SOC are two common concepts in battery health management and are usually used to describe the state of the battery. Among them, SOH represents the degree of attenuation of the current performance of the battery relative to the initial performance, reflecting the "lifespan state" of the battery. SOC represents the percentage of the current remaining charge of the battery to its full charge capacity, reflecting the "remaining charge" of the battery.

[0004] A neural network has a strong adaptability. It can learn from data and adapt to different patterns and trends, and is good at processing complex, non-linear, and large-scale data. It can process multiple inputs simultaneously, thus accelerating the processing speed. The neural network can also automatically learn and extract features from the input data without manual feature engineering. This makes the neural network more effective in processing complex, high-dimensional data and reduces the need for manual intervention. Summary of the Invention

[0005] In order to realize the monitoring of the prediction of the battery state of health, the present invention provides a method for jointly predicting battery SOH and SOC based on a neural network.

[0006] The present invention is realized through the following technical solutions: A method for jointly predicting the SOH and SOC of a battery based on a neural network, comprising the following steps:

[0007] Step (1), measure the AC impedance spectrum of the battery and record the battery charge and discharge data;

[0008] Step (2), import the data of the AC impedance spectrum into Z-view software for fitting to obtain the parameter values of the second-order RC equivalent circuit; use the Spearman rank correlation coefficient to screen the battery charge and discharge data;

[0009] Step (3), use the parameter values of the second-order RC equivalent circuit and the screened battery charge and discharge data as the input of the LSTM network;

[0010] Step (4), adopt the LSTM network (multi-scale convolutional network) to fully extract local features of the input of the battery state, use the LSTM network to extract the temporal features of the input time series, and set the dropout rate by the adaptive Dropout layer;

[0011] Step (5), the neuron connections obtained by training the LSTM network are fully connected layers, and the fully connected layers organize and output the prediction results of the battery SOC (remaining power) and SOH (life state).

[0012] As a further improvement of the technical solution of the present invention, in step (1), the measurement of the AC impedance spectrum of the battery is to measure the AC impedance spectrum of the battery by the lock-in amplifier method, specifically: perform operations on the excitation signal and the response signal flowing through the battery respectively with the reference signal, and obtain the real part and the imaginary part of the actual battery electrochemical impedance according to the real part and the imaginary part of the excitation signal and the response signal, and draw the electrochemical impedance spectrum graph; discard the high-frequency band, discard the parts with incorrect changes in the middle and low-frequency bands, and record the data of the AC impedance spectrum.

[0013] As a further improvement of the technical solution of the present invention, the method for obtaining the parameter values of the second-order RC equivalent circuit in step (2) specifically includes:

[0014] (2-1) Select the model: After importing the data of the AC impedance spectrum into Z-view software, select the second-order RC equivalent circuit for fitting;

[0015] (2-2) Set the initial parameters: Manually set the initial values of the resistance, capacitance and inductance components;

[0016] (2-3) Fitting adjustment: After setting the initial parameters, start the fitting process. Z-view software automatically adjusts the parameters of the second-order RC equivalent circuit to fit the model with the actual data. After the fitting is completed, manually adjust the parameters of the second-order RC equivalent circuit and evaluate the fitting quality.

[0017] As a further improvement of the technical solution of the present invention, the specific steps of using the Spearman rank correlation coefficient to screen the battery charge and discharge data in step (2) include:

[0018] Calculate the Spearman rank correlation coefficients between the battery charge and discharge data and the battery SOH and SOC. Select the data with a correlation coefficient greater than 0.8 as the input data of the LSTM network. Denoise the input data using the singular value decomposition method for the selected input data. Arrange the singular values in descending order and select the eigenvectors corresponding to the first half of the singular values.

[0019] As a further improvement of the technical solution of the present invention, step (4) specifically includes:

[0020] (4-1) Set the convolution kernel sizes of the convolution model to be 3, 4, and 5 respectively, and the activation function to be the RELU function;

[0021] (4-2) Use the adaptive pooling layer to reduce the dimension, and input it into the LSTM network after flattening through the Flatten layer;

[0022] (4-3) Input the time series into the LSTM network to extract the time series features of the time series;

[0023] (4-4) Introduce an adaptive Dropout layer, and use the activation function to calculate, and set the dropout rate according to the input specification of the neuron information.

[0024] As a further improvement of the technical solution of the present invention, step (5) specifically includes:

[0025] (5-1) Expand the neuron information output by the adaptive Dropout layer into a one-dimensional vector;

[0026] (5-2) Multiply the one-dimensional vector by the weight matrix of the fully connected layer;

[0027] (5-3) After multiplication, organize the neuron information through the fully connected layer and output the predicted values of the battery SOH and SOC.

[0028] The joint prediction method of battery SOH and SOC based on neural network provided by the present invention has the following advantages compared with the prior art:

[0029] 1) The present invention combines the alternating current impedance spectrum of the battery, the battery charge and discharge data with the LSTM network, imports the data into the Z-view software for fitting to obtain the parameter values of the second-order RC equivalent circuit, and uses the parameter values of the second-order RC equivalent circuit and the battery charge and discharge data as the input of the LSTM network, avoiding the insoluble errors that occur when only using neural networks.

[0030] (2) Use the Spearman rank correlation coefficient to screen highly correlated input data, reduce the training time and difficulty of the LSTM network, and use multi-scale convolutional kernels to extract local features; introduce an adaptive Dropout layer and specify the dropout rate.

[0031] (3) Set different methods according to different battery models in the charge and discharge experiments. Measure the electrochemical impedance spectrum with an electrochemical analyzer during the charge and discharge process, and fully integrate the battery mechanism with diversified inputs to make the prediction results more accurate. Description of the Drawings

[0032] The drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0034] Figure 1 It is the overall flowchart of the method for jointly predicting battery SOH and SOC based on neural network according to the present invention.

[0035] Figure 2 It is the fitting graph of the electrochemical impedance spectrum.

[0036] Figure 3 It is the second-order RC equivalent circuit.

[0037] Figure 4 It is the schematic diagram of the multi-scale convolutional neural network.

[0038] Figure 5 It is the schematic diagram of the relationship between the LSTM network, the Flatten layer, and the adaptive Dropout layer.

[0039] Figure 6 It is the schematic diagram of the connection relationship of the overall neural network of the present invention. Detailed Embodiments

[0040] In order to be able to more clearly understand the above objects, features, and advantages of the present invention, the following will further describe the solutions of the present invention. It should be noted that, without conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0041] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0042] The following details the specific embodiments of the present invention.

[0043] As Figure 1 shown, the present invention provides a specific embodiment of a method for jointly predicting the SOH and SOC of a battery based on a neural network, including the following steps:

[0044] Step (1), measure the AC impedance spectrum of the battery and record the battery charge and discharge data;

[0045] Step (2), import the data of the AC impedance spectrum into Z-view software for fitting to obtain the parameter values of the second-order RC equivalent circuit; use the Spearman rank correlation coefficient to screen the battery charge and discharge data;

[0046] Step (3), use the parameter values of the second-order RC equivalent circuit and the screened battery charge and discharge data as the input of the LSTM network;

[0047] Step (4), use the LSTM network to fully extract local features from the input of the battery state, extract the temporal features of the input time series using the LSTM network, and set the dropout rate by the adaptive Dropout layer;

[0048] Step (5), the neuron connections obtained by training the LSTM network are connected to the fully connected layer, and the fully connected layer organizes and outputs the prediction results of the battery SOH and SOC.

[0049] In step (1) of this embodiment, the measurement of the AC impedance spectrum of the battery is to measure the AC impedance spectrum of the battery using the lock-in amplifier method. The lock-in amplifier method is a common method for offline measuring the battery impedance spectrum, specifically:

[0050] 1) Measurement method: Operate the excitation signal and the response signal flowing through the battery respectively with the reference signal, and obtain the real part and the imaginary part of the actual battery electrochemical impedance according to the real part and the imaginary part of the excitation signal and the response signal, and draw the electrochemical impedance spectrum graph.

[0051] 2) Data processing: Since the high-frequency band in the electrochemical impedance spectrum graph is difficult to fit, it is discarded; as Figure 2 shown in the curve of the middle and low frequency bands in the fitting graph of the electrochemical impedance spectrum, since there is noise in the data of the middle and low frequency bands, the part with incorrect variation rules needs to be discarded; then record the data of the AC impedance spectrum.

[0052] In step (1) of this embodiment, when recording the battery charge and discharge data, the BMS is used to record the battery charge and discharge data. The battery charge and discharge data specifically includes: charge and discharge voltage, charge and discharge current, temperature, constant current charging time, equal voltage drop time, maximum discharge temperature point, time of minimum discharge voltage point, initial voltage sudden drop value, usage time, charge and discharge times, charge and discharge rate, etc.

[0053] The experimental method for battery charge and discharge adopted in this embodiment is as follows:

[0054] 1) Multi-step constant current and constant voltage charging:

[0055] 1-1) The multi-step constant current charging process is divided into three stages, and the charging rates are set to 5.6C, 4.3C, and 1C respectively;

[0056] 1-2) The charging rate is switched based on the battery SOC value, and the charging rate is switched at 54% and 80% respectively;

[0057] 1-3) When the voltage reaches 50% of the rated voltage, constant voltage charging is adopted until the rate reaches 0.02C and then the charging stops. The charging data of the battery needs to be recorded during the charging process.

[0058] 2) Constant current and constant voltage charging:

[0059] Charge at a charging rate of 0.75C. When the voltage reaches 50% of the rated voltage, constant voltage charging is carried out until the charging rate drops to 0.02C and then the charging stops. The charging data of the battery needs to be recorded during the charging process.

[0060] 3) Discharge process:

[0061] 3-1) The constant temperature conditions are: 0°C, 25°C, and 40°C respectively. Three temperatures are used to simulate the real conditions of battery use;

[0062] 3-2) Two-thirds of the batteries are connected to a resistive load machine for discharge, and the remaining batteries are connected to a DC motor for discharge. The discharge data of the battery needs to be recorded during the discharge process. The criterion for the end of one discharge is that the battery voltage drops to 20% of the rated voltage;

[0063] 3-3) The failure criterion for the battery cycle to stop is: the battery SOH becomes 80% of that at the time of leaving the factory.

[0064] In the above two charging methods, the multi-step constant current and constant voltage charging is for battery models No. 1 and No. 2. The constant current and constant voltage charging is for battery models No. 5 and No. 7. The discharge methods for different model batteries are the same.

[0065] When specifically applied, the method for obtaining the parameter values of the second-order RC equivalent circuit in step (2) specifically includes:

[0066] (2-1) Selection model: After importing the data of the electrochemical impedance spectroscopy into the Z-view software, select the second-order RC equivalent circuit for fitting; in this embodiment, the second-order RC equivalent circuit is as shown in Figure 3 shown, where represents the solution resistance, represents the charge transfer resistance, represents the SEI film resistance, represents the battery terminal voltage, represents the battery potential, CPE 1 and CPE 2 are constant phase elements respectively, which are often used to mimic the behavior of non-ideal electric double layer capacitors.

[0067] (2-2) Set initial parameters: Manually set the initial values of the resistors, capacitors and inductors in the second-order RC equivalent circuit.

[0068] (2-3) Fitting adjustment: After setting the initial parameters, start the fitting process. The Z-view software automatically adjusts the parameters of the second-order RC equivalent circuit to fit the model with the actual data. After the fitting is completed, manually adjust the parameters of the second-order RC equivalent circuit and evaluate the fitting quality.

[0069] In the step (2), regarding the screening of the battery charge and discharge data by using the Spearman rank correlation coefficient, specifically:

[0070] The calculation formula of the Spearman rank correlation coefficient is as follows:

[0071]

[0072] Where, represents the difference in sequence values between the th group of battery charge and discharge data, represents the number of groups of battery charge and discharge data.

[0073] Calculate the Spearman rank correlation coefficient between the battery charge and discharge data and the battery SOH and SOC , select the data with a correlation coefficient greater than 0.8 as the input data of the LSTM network. Denoise the selected input data by using the singular value decomposition method. Arrange the singular values in descending order, and select the eigenvectors corresponding to the first half of the singular values. These vectors have high correlations and can reduce the calculation difficulty. The calculation of the singular value decomposition method is as shown in the following formula:

[0074]

[0075] Where, represents 's input matrix, is The left singular matrix of is the right singular matrix of a diagonal matrix, represents the transpose, and the elements on the diagonal are the singular values of matrix arranged from largest to smallest, represents the number of groups of battery charge and discharge data, represents the number of each group of battery charge and discharge data, represents matrix rank of

[0076] Figure 4 is a schematic diagram of a multi-scale convolutional neural network. In this embodiment, the parameter values of the second-order RC equivalent circuit and the selected battery charge and discharge data are used as the input of the LSTM network. Assuming the input data is X i , j = [ x 11 ; x 12 ; x 13 ; ⋅⋅⋅ ; x nm ] ∈ ℝ n × m , its convolution is defined as follows:

[0077]

[0078] where is the feature map after convolution calculation using the th convolution kernel and the input data, and are the weights and biases of the th convolution kernel respectively, represents the convolution kernel size, and represent the th row and the th column in the input data respectively.

[0079] Step (4) specifically includes:

[0080] (4-1) Set the convolution kernel sizes of the convolution model to 3, 4, and 5 respectively, and the activation function to the RELU function.

[0081] Among them, the RELU function (rectified linear unit function) is as follows, which can perform a non-linear transformation on the feature map to obtain the features extracted by the convolution layer.

[0082]

[0083] where represents the calculated value of the RELU function, represents the function symbol.

[0084] (4-2) Use the adaptive pooling layer to reduce the dimension and filter out invalid information from the scaled mapping of the features; since the convolution kernel sizes are different, the feature sizes extracted by the convolutional layer are also different. To facilitate input into the LSTM network, use the Flatten layer to flatten the features.

[0085] (4-3)Input the time series into the LSTM network to extract the time series features, which can effectively capture and utilize the long-term dependencies in the time series;

[0086] (4-4)Introduce an adaptive Dropout layer and use the activation function to calculate and set the dropout rate according to the input specification of the neuron information.

[0087] During the calculation process of the LSTM network, overfitting is likely to occur. Adaptive Dropout is an improved Dropout technique. Therefore, in this embodiment, by introducing the adaptive Dropout layer, an activation function is introduced, as shown below:

[0088]

[0089] This formula dynamically adjusts the dropout probability according to the input of the neuron information, thereby determining whether a neuron should be dropped. In the formula, represents the probability of the th neuron being retained, represents the participation degree of the th neuron, represents the number of neurons.

[0090] Step (5) of this embodiment specifically includes:

[0091] (5-1)Expand the neuron information output by the adaptive Dropout layer into a one-dimensional vector;

[0092] (5-2)Multiply the one-dimensional vector by the weight matrix of the fully connected layer;

[0093] (5-3)After multiplication, the fully connected layer organizes the neuron information and outputs the predicted values of the battery SOH and SOC.

[0094] To verify the prediction accuracy of the battery SOH and SOC in the battery SOH and SOC joint prediction method based on neural network described in this embodiment, the following verification method is adopted:

[0095] Use the Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Square Error (MSE) as evaluation criteria to verify the prediction results:

[0096] The experimental results show that: when the MAPE value is between 10% and 20%, the MAE value is between 1 and 4, the RMSE value is between 1 and 2, the MSE value is between 1 and 4, and the MAE value is between 1 and 5, it indicates that the prediction method of this embodiment has a relatively high accuracy.

[0097] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Although the foregoing embodiments have been described in detail, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A battery SOH and SOC joint prediction method based on neural network, characterized in that: The steps include: Step (1), measuring the AC impedance spectrum of the battery and recording the battery charge and discharge data; Step (2), importing the data of the AC impedance spectrum into the Z-view software for fitting to obtain the parameter values ​​of the second-order RC equivalent circuit; using the Spearman rank correlation coefficient to screen the battery charge and discharge data; Step (3), using the parameter values ​​of the second-order RC equivalent circuit and the screened battery charge and discharge data as inputs of the LSTM network; Step (4), using the LSTM network to fully extract local features of the battery state input, using the LSTM network to extract the timing features of the input timing sequence, and using the adaptive Dropout layer to set the dropout rate; Step (5), the neurons obtained through LSTM network training are connected to the fully connected layer, and the fully connected layer organizes and outputs the battery SOH and SOC prediction results.

2. The method for joint prediction of battery SOH and SOC based on neural network according to claim 1, characterized in that: The AC impedance spectrum of the battery is measured in step (1) by using a lock-in amplifier method, specifically: the excitation signal and the response signal flowing through the battery are respectively operated with the reference signal, and the real part and the imaginary part of the actual electrochemical impedance of the battery are obtained according to the real part and the imaginary part of the excitation signal and the response signal, and the electrochemical impedance spectrum graph is drawn; Discard the high frequency band, discard the incorrect changes in the mid and low frequency bands, and record the data of the AC impedance spectrum.

3. The method for joint prediction of battery SOH and SOC based on neural network according to claim 1, characterized in that: The method for obtaining the parameter values ​​of the second-order RC equivalent circuit in step (2) specifically includes: (2-1) Model selection: After importing the AC impedance spectrum data into the Z-view software, select the second-order RC equivalent circuit for fitting; (2-2) Set initial parameters: manually set the initial values ​​of resistance, capacitance and inductance components; (2-3) Fitting adjustment: After setting the initial parameters, start the fitting process. The Z-view software automatically adjusts the parameters of the second-order RC equivalent circuit to make the model fit the actual data. After the fitting is completed, manually adjust the parameters of the second-order RC equivalent circuit and evaluate the fitting quality.

4. The method for jointly predicting battery SOH and SOC based on a neural network according to claim 1, characterized in that: The specific steps of using the Spearman rank correlation coefficient to screen the battery charge and discharge data in step (2) include: The Spearman rank correlation coefficient between the battery charge and discharge data and the battery SOH and SOC is calculated, and the data with a correlation coefficient greater than 0.8 is selected as the input data of the LSTM network. The selected input data is denoised using the singular value decomposition method, the singular values ​​are arranged in descending order, and the eigenvectors corresponding to the first half of the singular values ​​are selected.

5. The method for joint prediction of battery SOH and SOC based on neural network according to claim 1, characterized in that: The step (4) specifically includes: (4-1) Set the convolution kernel sizes of the convolution model to 3, 4, and 5 respectively, and the activation function to the RELU function; (4-2) Use the adaptive pooling layer to reduce the dimension, and then input it into the LSTM network after flattening through the Flatten layer; (4-3) Input the time series sequence into the LSTM network to extract the time series features of the time series sequence; (4-4) Introduce the adaptive Dropout layer and use The activation function is calculated and the dropout rate is set based on the input specification of the neuron information.

6. A battery SOH and SOC joint prediction method based on neural network according to claim 5, characterized in that: The step (5) specifically includes: (5-1) Expand the neuron information output by the adaptive Dropout layer into a one-dimensional vector; (5-2) Multiply the one-dimensional vector by the weight matrix of the fully connected layer; (5-3) After multiplication, the neuron information is sorted out through the fully connected layer and the battery SOH and SOC prediction values ​​are output.

Images