A lithium-ion battery SOC estimation method based on fusion neural network

Abstract

The present invention provides a lithium-ion battery SOC estimation method based on a fusion neural network, which belongs to the field of lithium-ion battery technology. It solves the technical problem that the current battery capacity SOC estimation is difficult to estimate. The technical solution is: step 1) fully charge a new lithium battery and measure the terminal voltage, current, and temperature of the lithium-ion battery when the SOC ranges from 1 to 0; step 2) pre-process the measured data; step 3) use a 2DCNN convolutional neural network to train and test the new data set to achieve real-time SOC estimation. The beneficial effects of the present invention are: the present invention adds estimated battery capacity data to the data set, which can have higher estimation accuracy, and compared with LSTM, it will not always remember the error; and the short historical data length can also have a faster estimation speed; compared with using 3DCNN to estimate SOC, it can reduce the amount of calculation and reduce unnecessary error input, thereby having higher accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a lithium-ion battery SOC estimation method based on a fusion neural network. Background Art

[0002] Electric vehicles have become the current trend in the automotive industry, and lithium batteries are widely used in electric vehicles due to their many advantages. In automotive battery management systems, accurate state of charge (SOC) estimation is crucial for calculating remaining range and helps the management system more effectively manage and control the vehicle battery.

[0003] Currently, there are three main types of SOC estimation methods. The first is the ampere-hour integration method. This method requires a clear initial SOC value, making it difficult to apply in practice, but it is mostly used in SOC definition. The second type is model-based methods, mainly including electrochemical models and equivalent circuit methods. Due to their complex structure and excessive number of parameters, electrochemical models are not very popular for mobile phone or automotive batteries. Equivalent circuit model-based methods, which include both online and offline methods, identify parameters and then combine filtering algorithms to estimate SOC. While highly accurate, they struggle to account for changes in temperature and aging over time. The third type is data-driven methods, mainly including support vector machines and neural networks. Data-driven models do not require specific models and can comprehensively consider multiple factors, resulting in higher accuracy and robustness. In practical applications, we often need to obtain data such as the remaining charge in mobile phone and automotive batteries, which requires careful consideration of factors such as temperature and aging. Battery capacity is not only an indicator of aging but also a key parameter in defining SOC values. Therefore, SOC estimation based on battery capacity is a technical challenge.

[0004] How to solve the above technical problems is the subject faced by the present invention. Summary of the Invention

[0005] The purpose of the present invention is to provide a lithium-ion battery SOC estimation method based on a fusion neural network. The battery parameters are obtained through battery charge and discharge experiments at different temperatures. The data is then processed and a training data set is constructed. The trained data set is imported into a 3DCNN convolutional neural network for training. After the battery capacity is estimated, the estimated battery capacity is added to the data set to construct a new data set, which can ultimately be used for real-time estimation of the lithium battery SOC.

[0006] The present invention is achieved by the following measures: a lithium-ion battery SOC estimation method based on a fusion neural network, comprising the following steps:

[0007] Step 1) Fully charge a new lithium-ion battery under varying temperatures, then repeatedly measure the terminal voltage, current, and temperature of the lithium-ion battery as the SOC changes from 1 to 0 through a constant current discharge test, a DST discharge test, a FUDS discharge test, and a US06 discharge test;

[0008] Step 2) Preprocessing the measured data to construct a 3DCNN training dataset and a test dataset for estimating battery capacity, and constructing a new dataset using the battery capacity estimated by the 3DCNN convolutional neural network;

[0009] Step 3) Use the 2DCNN convolutional neural network to train and test the new dataset to achieve the final SOC real-time estimation.

[0010] As a further optimization scheme of the lithium-ion battery SOC estimation method based on a fusion neural network provided by the present invention, the step 2) specifically includes the following steps:

[0011] Step 2-1) The lack of uniform data dimensions will lead to slow convergence and high errors when fitting the neural network algorithm, so the measured data needs to be normalized. The battery objects of various working condition experiments are the same, but the range of variation of voltage and current is different. Therefore, the data of various working conditions need to be pre-processed separately first. The following formula is used to normalize the input data:

[0012]

[0013] Among them, a ij is the jth data point of the i-th input, a min is the minimum value of the i-th input data, a max is the maximum value of the i-th input data, x inij is a ij The data after normalization.

[0014] The output is the SOC of the lithium battery. The SOC range is between 0 and 1, so normalization is not required.

[0015] Step 2-2) After data normalization, in order to explore the relationship between the input data in a discharge cycle and the relationship between adjacent discharge cycles, the following method is used to construct a data set. First, the measurement values ​​in a discharge cycle are truncated into several data of length n, as shown below:

[0016]

[0017] Among them, x ijIndicates the jth frame data of the i-th input in the input data set. Since the 3DCNN convolutional neural network can simultaneously input multiple frames of data at a timestamp, the voltage and current data of j consecutive discharge cycles in the same discharge time period are selected as an input of the 3DCNN. In the present invention, j is 3. In the data set, T represents the temperature of the battery during operation, U represents the voltage at both ends of the battery, I represents the current passing through the battery, and t represents the discharge time. In the present invention, n is 10, then x ij The corresponding output y i t n The battery capacity value at the third frame.

[0018] Step 2-3) For a discharge experiment with a total number of charge and discharge cycles N and a single discharge cycle t = m, we use the padding method to expand the edge data to fully explore the connections between the input data before and after each discharge cycle and the connections between the input data between adjacent discharge cycles. Then, we generate a total of N × m groups of input data of size 3 × 4 × 10, and the corresponding N × m SOC values ​​are output data. The input and output data of different operating conditions are combined into a total data set.

[0019] In steps 2-4, the training and test datasets were extracted without shuffling the dataset to ensure that the training data would not be affected. The first 80% of the data from a discharge condition was used as the training dataset, and the last 20% of the data was used as the test dataset. These were used as training and test data for the 3DCNN.

[0020] Step 2-5) The 3DCNN convolutional neural network architecture developed for lithium-ion battery capacity estimation includes: an input layer (Input), three convolutional layers (C), two fully connected layers (F) and an output layer (Output).

[0021] Step 2-6) All convolutional layers and fully connected layers are activated using the ReLU function, and the output layer is activated using the Tanh function, and mapped as follows.

[0022] Step 2-7) Add the estimation of battery capacity and construct a new input data set with a size of 5×10 and an output data set of t n The SOC value corresponding to the moment.

[0023] As a further optimization solution of the lithium-ion battery SOC estimation method based on a fusion neural network provided by the present invention, step 3) specifically includes the following steps:

[0024] Step 3-1) The 2DCNN convolutional neural network architecture developed for lithium-ion battery SOC estimation includes: an input layer, three convolutional layers, two fully connected layers, and an output layer.

[0025] Step 3-2) The convolutional layers and fully connected layers are all activated using the relu function, and the output layer is activated using the tanh function and mapped.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) The present invention uses a fusion neural network algorithm based on battery capacity, which adds estimated battery capacity data to the data set and can achieve higher estimation accuracy.

[0028] (2) The present invention uses historical data with a length of 10 or less, and can memorize the historical data of previous cycles, with long-term and short-term memory capabilities. Compared with LSTM, it does not permanently memorize errors; and the short historical data length also enables faster estimation speed.

[0029] (3) The 2DCNN used in the present invention to estimate SOC can reduce the amount of computation and unnecessary error input compared to using 3DCNN to estimate SOC, according to the definition of SOC, thereby achieving higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0031] Figure 1 It is the overall framework flow chart of the present invention.

[0032] Figure 2 Schematic diagram of the new dataset construction method in the present invention.

[0033] Figure 3 This is a curve diagram of the estimated battery capacity in the present invention.

[0034] Figure 4 This is a schematic diagram of the SOC estimation results of the first cycle of the test set in Example 1 of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0036] Example 1

[0037] See also Figures 1 to 4The present invention provides a method for estimating the state of charge (SOC) of a lithium-ion battery based on a fused neural network. This embodiment utilizes a Panasonic NCR-18650B lithium-ion battery with a nominal voltage of 3.7V and a capacity of 3400mAh. The battery is fully charged using a constant current / constant voltage charging method. After standing for 1 hour, the battery is fully charged. Discharge tests are then repeated using constant current discharge, DST, FUDS, and US06 conditions until the voltage drops to the discharge cutoff voltage.

[0038] In order to better achieve the purpose of the present invention, this embodiment is a lithium-ion battery SOC estimation method based on a fusion neural network, comprising the following steps:

[0039] Step 1) Fully charge a new lithium-ion battery under varying temperatures, then repeatedly measure the terminal voltage, current, and temperature of the lithium-ion battery as the SOC changes from 1 to 0 through a constant current discharge test, a DST discharge test, a FUDS discharge test, and a US06 discharge test;

[0040] Step 2) Preprocessing the measured data to construct a 3DCNN training dataset and a test dataset for estimating battery capacity, and constructing a new dataset using the battery capacity estimated by the 3DCNN convolutional neural network;

[0041] Step 3) Use the 2DCNN convolutional neural network to train and test the new dataset to achieve the final SOC real-time estimation.

[0042] Preferably, the step 2) specifically includes the following steps:

[0043] Step 2-1) The lack of uniform data dimensions will lead to slow convergence and high errors when fitting the neural network algorithm, so the measured data needs to be normalized. The battery objects of various working condition experiments are the same, but the range of variation of voltage and current is different. Therefore, the data of various working conditions need to be pre-processed separately first. The following formula is used to normalize the input data:

[0044]

[0045] Among them, a ij is the jth data point of the i-th input, a min is the minimum value of the i-th input data, a max is the maximum value of the i-th input data, x inij is a ij The data after normalization.

[0046] The output is the SOC of the lithium battery. The SOC range is between 0 and 1, so normalization is not required.

[0047] Step 2-2) After data normalization, in order to explore the relationship between the input data in a discharge cycle and the relationship between adjacent discharge cycles, the following method is used to construct a data set. First, the measurement values ​​in a discharge cycle are truncated into several data of length n, as shown below:

[0048]

[0049] Among them, x ij Indicates the jth frame data of the i-th input in the input data set. Since the 3DCNN convolutional neural network can simultaneously input multiple frames of data at a timestamp, the voltage and current data of j consecutive discharge cycles in the same discharge time period are selected as an input of the 3DCNN. In the present invention, j is 3. In the data set, T represents the temperature of the battery during operation, U represents the voltage at both ends of the battery, I represents the current passing through the battery, and t represents the discharge time. In the present invention, n is 10, then x ij The corresponding output y i t n The battery capacity value at the third frame.

[0050] Step 2-3) For a discharge experiment with a total number of charge and discharge cycles N and a single discharge cycle t = m, we use the padding method to expand the edge data to fully explore the connections between the input data before and after each discharge cycle and the connections between the input data between adjacent discharge cycles. Then, we generate a total of N × m groups of input data of size 3 × 4 × 10, and the corresponding N × m SOC values ​​are output data. The input and output data of different operating conditions are combined into a total data set.

[0051] In steps 2-4, the training and test datasets were extracted without shuffling the dataset to ensure that the training data would not be affected. The first 80% of the data from a discharge condition was used as the training dataset, and the last 20% of the data was used as the test dataset. These were used as training and test data for the 3DCNN.

[0052] Step 2-5) The 3DCNN convolutional neural network architecture developed for lithium-ion battery capacity estimation includes: an input layer (Input), three convolutional layers (C), two fully connected layers (F), and an output layer (Output). Specifically: in the convolutional layer, the first two layers are 32 three-dimensional convolution kernels of 2×2×2, and the third layer is 64 two-dimensional convolution kernels of 2×2; the number of fully connected layers is 64 and 32, and the estimation results are as follows: Figure 3 shown.

[0053] Step 2-6) The convolutional layers and fully connected layers are all activated using the relu function, and the output layer is activated using the tanh function and mapped.

[0054] Step 2-7) Add the estimation of battery capacity and construct a new input data set. The size of the new input data set is 5×10 and the output data set is t n The SOC value corresponding to the moment.

[0055] Preferably, the step 3) specifically comprises the following steps:

[0056] Step 3-1) The 2DCNN convolutional neural network architecture developed for lithium-ion battery SOC estimation includes: an input layer, three convolutional layers, two fully connected layers, and an output layer. Specifically: in the convolutional layer, three layers are 32 two-dimensional convolution kernels of 2×2; the number of fully connected layers is 64 and 32. The estimation result of a certain cycle is as follows Figure 4 shown.

[0057] Step 3-2) The convolutional layers and fully connected layers are all activated using the relu function, and the output layer is activated using the tanh function and mapped.

[0058] A battery capacity-based fusion neural network algorithm used in an embodiment of the present invention adds estimated battery capacity data to a data set, which can achieve higher estimation accuracy.

[0059] The historical data used in this embodiment is 10 or less in length, and it can retain the historical data of previous cycles, possessing long-term and short-term memory capabilities. Compared to LSTM, it does not permanently retain errors; and the short historical data length also enables faster estimation speed.

[0060] The 2DCNN used in the embodiment of the present invention to estimate the SOC can reduce the amount of computation and unnecessary error input compared to using 3DCNN to estimate the SOC, according to the definition of SOC, thereby achieving higher accuracy.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A lithium-ion battery SOC estimation method based on a fusion neural network, characterized in that: The following steps are involved: Step 1) Fully charge a new lithium-ion battery under varying temperatures, then repeatedly measure the terminal voltage, current, and temperature of the lithium-ion battery as the SOC changes from 1 to 0 through a constant current discharge test, a DST discharge test, a FUDS discharge test, and a US06 discharge test; Step 2) Preprocessing the measured data to construct a 3DCNN training dataset and a test dataset for estimating battery capacity, and constructing a new dataset using the battery capacity estimated by the 3DCNN convolutional neural network; The step 2) specifically includes the following steps: Step 2-1) Normalize the measured data. The battery objects of various working condition experiments are the same, but the voltage and current variation ranges are different. First, pre-process the data of various working conditions separately and use the following formula to normalize the input data: Among them, a ij is the jth data point of the i-th input, a min is the minimum value of the i-th input data, a max is the maximum value of the i-th input data, x inij is a ij The data after normalization; The output is the SOC of the lithium battery. The SOC range is between 0 and 1 and does not require normalization. Step 2-2) After data normalization, we explore the connections between the input data in a discharge cycle and the connections between adjacent discharge cycles. We construct a data set using the following method: First, we truncate the measured values ​​in a discharge cycle into several data of length n, as shown below: Among them, x ij Represents the j-th frame data of the i-th input in the input data set. The 3DCNN convolutional neural network simultaneously inputs multiple frames of data at a timestamp. By selecting the voltage and current data of j consecutive discharge cycles in the same discharge time period as an input of the 3DCNN, in the data set, T represents the temperature of the battery during operation, U represents the voltage across the battery, I represents the current passing through the battery, and t represents the discharge time. Then x ij The corresponding output y i t n The battery capacity value at the 3rd frame; Step 2-3) For a discharge experiment with a total number of charge and discharge cycles N and a single discharge cycle t = m, in order to explore the relationship between the input data before and after each discharge cycle and the relationship between the input data between adjacent discharge cycles, the edge data is expanded using the padding method. A total of N × m groups of input data with a size of 3 × 4 × 10 are generated, and the corresponding N × m SOC values ​​are output data. The input and output data of different working conditions are merged into the total data set; Step 2-4) Select the first 80% of the data of a discharge condition as the training data set, and the last 20% of the data as the test data set, which are used as training and test data for 3DCNN; Steps 2-5) The 3DCNN convolutional neural network architecture developed for lithium-ion battery capacity estimation includes: an input layer (Input), three convolutional layers (C), two fully connected layers (F), and an output layer (Output); Step 2-6) All convolutional layers and fully connected layers are activated using the relu function, and the output layer is activated using the tanh function and mapped; Step 2-7) Add the estimation of battery capacity and construct a new input data set with a size of 5×10 and an output data set of t n SOC value corresponding to the moment; Step 3) Use the 2DCNN convolutional neural network to train and test the new dataset to achieve the final SOC real-time estimation; The step 3) specifically comprises the following steps: Step 3-1) The 2DCNN convolutional neural network architecture developed for lithium-ion battery SOC estimation includes: an input layer, three convolutional layers, two fully connected layers, and an output layer; Step 3-2) The convolutional layers and fully connected layers are all activated using the relu function, and the output layer is activated using the tanh function and mapped.

Images