CN116224074A - A method, device, and storage medium for estimating the state of charge of a soft-pack lithium-ion battery - Google Patents

Abstract

The invention relates to a soft package lithium ion battery state of charge estimation method, a device and a storage medium based on dynamic stress and deep learning, wherein the method comprises the following steps: establishing a battery charge-discharge data set containing dynamic stress signals, wherein the battery charge-discharge data set comprises current, terminal voltage, dynamic stress data and a reference state of charge; offline training is based on dynamic stress and deep learning soft package lithium ion battery state of charge estimation model: according to the input form required by the deep learning network, recombining the time sequence characteristics of current, terminal voltage and dynamic stress in a battery charge and discharge data set into a three-dimensional tensor sample set, constructing a deep learning network structure of a battery state-of-charge estimation model, and offline training the battery state-of-charge estimation model based on a reference state-of-charge and the sample set; on-line battery state of charge estimation. Compared with the prior art, the method has the advantages of strong adaptability to different working conditions, good estimation performance when the data length is limited, strong robustness to measurement noise interference and the like.

Description

A method, device, and storage medium for estimating the state of charge of a soft-pack lithium-ion battery

technical field

The invention relates to the field of electric vehicle energy storage batteries, in particular to a method, device and storage medium for estimating the state of charge of a soft-pack lithium-ion battery based on dynamic stress and deep learning.

Background technique

At present, the low-carbon vision promotes the development of new energy vehicles, and one of the important routes is pure electric vehicles. The power battery is responsible for energy storage and power output. Due to the advantages of high energy/power density, long service life and environmental friendliness, lithium-ion batteries have become a research hotspot for vehicle manufacturers, universities and scientific research institutions. The state of charge of the battery is defined as the ratio of the remaining power to the maximum available capacity, which is used to indicate the remaining power of the battery and is an important state quantity to determine the remaining mileage, charging/discharging power, and safe working range.

Conventional SOC estimation methods include ampere-hour integration methods, model-based methods, and data-driven methods. The ampere-hour integration method is based on the definition of the state of charge, which is an open-loop estimation, which is greatly affected by the initial error and measurement noise. The model-based method estimates the state of charge of the battery by establishing a battery model and constructing an adaptive filter. The estimation accuracy, adaptability and anti-interference ability to measurement noise are greatly affected by the model and the observer. The data-driven approach implements state-of-charge estimation based on the relationship between battery measurable information and internal state.

The general idea of the data-driven method is to measure the current, voltage, temperature, etc. of the battery when the battery is charging and discharging, use these data directly as input, build a data-driven model and train it to obtain appropriate model parameters, and use the trained model to perform battery charging. Electrical state estimation. Therefore, there are two main issues that need to be considered to improve the estimation performance of data-driven methods: one is how to select and build a strong adaptable data-driven model, and the other is how to obtain high-quality battery data.

The current data-driven models commonly used for SOC estimation are mainly divided into traditional machine learning models and neural network models. Traditional machine learning methods include Gaussian process regression and support vector machines; neural network models include convolutional neural networks and recurrent neural networks. Compared with traditional machine learning techniques, deep learning techniques can handle relatively larger sliding window lengths, and deep learning models for processing time series include long short-term memory neural networks and gated recurrent neural networks.

Battery sensing technology and multi-dimensional data are the key to data-driven battery state estimation. Conventional state-of-charge estimation methods generally use electrical signals and thermal signals to achieve state-of-charge estimation. Current state-of-charge estimation based on data-driven methods usually takes measured values such as battery current, voltage, and temperature within a sliding window as model input. With the development of advanced battery sensing technology, it is expected to utilize more sensing signals for battery state estimation, such as battery stress signals. During charging and discharging of the battery, the intercalation and deintercalation of lithium ions from the electrode active material will lead to structural changes in the electrode material, which in turn will cause changes in the volume of the battery, resulting in stress changes.

Combining advanced sensing techniques and machine learning techniques holds promise for more refined battery state estimation. At present, the main problems of this method are: first, the current data-driven model does not consider the time series characteristics of battery measurement data, and cannot effectively use historical information; Accurate, efficient and reliable battery SOC estimation.

Contents of the invention

The purpose of the present invention is to provide a method, device, and storage medium for estimating the state of charge of a soft-pack lithium-ion battery based on dynamic stress and deep learning, taking into account the time series characteristics of battery measurement data, effectively using the stress information of the battery, and improving the battery life. Reliability of state of charge estimation.

The purpose of the present invention can be achieved through the following technical solutions:

A method for estimating the state of charge of a soft-pack lithium-ion battery based on dynamic stress and deep learning, comprising the following steps:

Step 1) Establishing a battery charge and discharge data set containing dynamic stress signals, said battery charge and discharge data set including current, terminal voltage, dynamic stress data and reference state of charge;

Step 2) off-line training based on dynamic stress and deep learning soft pack lithium-ion battery state of charge estimation model;

Step 21) reorganize the time series features of the current, terminal voltage and dynamic stress in the battery charge and discharge data set into a three-dimensional tensor sample set according to the required input form of the deep learning network;

Step 22) Build the deep learning network structure of the battery state of charge estimation model, and train the battery state of charge estimation model offline based on the reference state of charge and sample set;

Step 3) Online battery state of charge estimation:

Step 31) Obtaining battery current, terminal voltage and dynamic stress signals online during the actual charging and discharging process of the battery;

Step 32) Perform online estimation of the state of charge of the power battery according to the trained battery state of charge estimation model.

Described step 1) comprises the following steps:

Step 11) Build a pouch battery stress measurement experimental device by using a restraint fixture and a strain sensor;

Step 12) Design battery charging and discharging experiment according to the ambient temperature and charging and discharging current recommended by the battery, test according to the set working conditions, record the battery current, terminal voltage and dynamic stress during the test, and calculate the reference state of charge of the battery, Build a battery charge and discharge dataset.

In the step 11), the plane stress measurement is carried out through the restraint fixture and the stress sensor in the laboratory environment. In the vehicle application, the stress sensor is arranged between the two pouch batteries to measure the plane stress.

In the step 12), the dynamic stress S _d recorded during the test is the stress related to the battery operating state, which is the value after subtracting the initial static stress S _s from the total stress S _t :

S _d =S _t -S _s

Wherein, the total stress is the instantaneous stress measured directly by the pressure sensor on the surface of the battery during charging/discharging, and the static stress is the stress obtained after the battery reaches an equilibrium state with sufficient rest time;

The battery reference state of charge is calculated by definition:

Among them, SOC _k and I _k are the state of charge and current of the battery at time k, respectively, Δt is the sampling time, and C _m is the battery capacity.

In the step 12), the working conditions set include: long-term charging and discharging working conditions, short-term charging and discharging working conditions, pulse charging and discharging working conditions and dynamic driving working conditions.

In the step 21), the battery charge and discharge data at time k is x _k =[I _k , V _k , F _k ], where I _k is the current at time k, V _k is the terminal voltage at time k, and F _k is the dynamic stress at time k;

Using a sliding window with a length of n, the input for training the deep learning network is X _k = [x _k-n+1 ,...,x _k-1 , x _k ], and the data length is recorded as N _data , then the obtained The sample size is:

N _samples =N _data -n+1

Reorganize the data into a three-dimensional tensor form: [N _samples ,n,N _features ], where N _features represent the number of features.

In said step 22), the deep learning network includes one layer of input layer, two layers of LSTM hidden layers, one layer of fully connected layer and one layer of output layer, wherein the LSTM network activation process is as follows:

是更新过程的候选状态，W_x是权重，b_x是偏置，σ(·)是sigmoid激活函数，tanh(·)是双曲正切激活函数。Among them, _ft forget gate, _it is the input gate, o _t is the output gate,

is the candidate state for the update process, W _x is the weight, b _x is the bias, σ( ) is the sigmoid activation function, and tanh( ) is the hyperbolic tangent activation function.

In the step 32), the method for estimating the state of charge of the battery is:

SOC ^* ＝LSTM(x ^* )

Among them, LSTM is a trained battery state of charge estimation model, and x ^* is an input vector composed of battery current, terminal voltage and dynamic stress obtained online.

A plug-and-charge function test device for a charging pile, comprising a memory, a processor, and a program stored in the memory, and the processor implements the above-mentioned method when executing the program.

A storage medium, on which a program is stored, and when the program is executed, the method as described above is implemented.

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

The method for estimating the state of charge of the battery in the present invention considers the dynamic stress signal in addition to the current and voltage, and can reflect the volume change caused by the insertion and extraction of lithium ions during the charging and discharging process of the battery. Combined with the deep learning network, the online estimation of the state of charge of the battery is realized. The model has strong adaptability to different working conditions, good estimation performance in the case of limited data length, and strong robustness to battery measurement noise interference.

Description of drawings

Fig. 1 is method flowchart of the present invention;

2 is a schematic diagram of the relationship between battery stress and state of charge;

Figure 3 shows the state of charge estimation results of soft-pack lithium-ion batteries based on dynamic stress and deep learning;

Fig. 4 is a diagram of the state of charge estimation error estimated based on the method of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

This embodiment provides a method for estimating the state of charge of a soft-pack lithium-ion battery based on dynamic stress and deep learning, as shown in FIG. 1 , including the following steps:

Step 1) Establish a battery charge and discharge data set containing dynamic stress signals, the battery charge and discharge data set includes current, terminal voltage, dynamic stress data and reference state of charge.

Step 11) Build a pouch battery stress measurement experimental device by using a restraint fixture and a strain sensor.

In addition to measuring voltage, current and temperature during battery operation, it is necessary to measure the plane stress of the battery. In the laboratory environment, the plane stress measurement is carried out through the restraint fixture and the stress sensor. In the vehicle application, the stress sensor is arranged between two pouch batteries to measure the plane stress.

The battery used in this implementation case is a soft-pack lithium-ion battery, and the positive and negative electrode materials are lithium manganate and graphite respectively. The nominal capacity of this battery is 8Ah, the nominal voltage is 3.7V, the charge cut-off voltage is 4.2V, and the discharge cut-off voltage is 2.8V.

Step 12) Design battery charging and discharging experiment according to the ambient temperature and charging and discharging current recommended by the battery, test according to the set working conditions, record the battery current, terminal voltage and dynamic stress during the test, and calculate the reference state of charge of the battery, Build a battery charge and discharge dataset.

The intercalation and deintercalation of lithium ions between the two electrodes during battery charging and discharging will cause the volume change of the battery and cause stress. The dynamic stress S _d recorded during the test is the stress related to the operating state of the battery (such as current, state of charge, etc.), which is the value after subtracting the initial static stress S _s from the total stress S _t :

S _d =S _t -S _s

Among them, the total stress is the instantaneous stress measured directly by the pressure sensor on the battery surface during charging/discharging, and the static stress is the stress obtained after a sufficient rest period (usually 2 hours) to bring the battery to an equilibrium state.

In this step, the working conditions set should cover the actual working conditions of the battery as much as possible. Improve the fitness of the estimated model.

In this embodiment, four typical working conditions are designed to simulate the actual charge and discharge of the battery.

Working condition 1: Constant current working condition, simulating the long-term charge and discharge of the battery. Charge and discharge at a constant current rate of 1C, denoted as Cy1 and Cy2.

Working condition 2: short-term working condition, simulating short-term charge and discharge of the battery. The batteries were charged to a set state of charge at different current rates (for example, 0.25C, 0.5C, 1.5C, and 2C), and then discharged to 0% state of charge at a current of 1C, denoted as Cy3-Cy6, respectively. The set states of charge are 20%, 40%, 60% and 80%, respectively.

Working condition three: pulse working condition, simulating the power characteristics of the battery. This working condition includes two sub-working conditions: Sub-working condition 1 uses different current ratios (0.25C, 0.5C, 1.5C and 2C) to charge the battery to the set state of charge, then let the battery stand for 2 hours, record It is Cy7-Cy10; sub-condition two is firstly charged to the set state of charge with a constant current of 1C, and then loaded with 60-second discharge-charge pulses of different current rates (0.25C, 0.5C, 1.5C and 2C). Stand still for 1 hour between two adjacent charge state points, which are respectively recorded as Cy11-Cy14.

Working condition four: dynamic working condition, simulating the actual driving of electric vehicles. Two commonly used driving conditions - the New European Driving Cycle (NEDC) and the Urban Road Cycle (UDDS), denoted as Cy15 and Cy16 respectively.

The measured cell strain is the strain value after subtracting the initial stress. During the experiment, the fixture and test cells were placed in an incubator to ensure environmental consistency, and the ambient temperature was set at 25 °C.

According to the definition, the reference state of charge of the battery is calculated by current integration, and the reference state of charge is used as the output label of offline training. The state of charge of the battery is defined as the ratio of the remaining usable power of the battery to the current capacity of the battery, which can be calculated by adding the SOC _k-1 of the previous moment to the current changing power I _k Δt (I _{k is negative when discharging, and I k} is negative when charging. _k is positive) and the capacity C _m is calculated as a ratio. In this embodiment, the sampling time Δt=1s, and the battery capacity C _m =8Ah. The battery reference state of charge is then calculated as:

As shown in Figure 2, the relationship between battery stress and state of charge is less affected by the battery operating current. The relationship between battery stress/strain and state of charge is quantitatively evaluated by Pearson and Spearman correlation coefficients. Pearson correlation coefficient indicates that dynamic stress is linearly related to state of charge, and Spearman correlation coefficient indicates a monotonic relationship between dynamic stress and state of charge. The two correlation coefficients between battery dynamic stress and state of charge exceed 0.97 and 0.985, respectively, indicating a strong correlation between the battery dynamic stress signal and state of charge, suitable for use as an input to the state of charge estimation model.

Step 2) Offline training of the state-of-charge estimation model for soft-pack lithium-ion batteries based on dynamic stress and deep learning.

Step 21) Reorganize the time series features of current, terminal voltage and dynamic stress in the battery charge and discharge data set into a three-dimensional tensor sample set according to the required input form of the deep learning network.

The battery charge and discharge data at time k is x _k = [I _k , V _k , F _k ], where I _k is the current at time k, V _k is the terminal voltage at time k, and F _k is the dynamic stress at time k.

Using a sliding window with a length of n, that is, using the battery sensor data from time k-n+1 to k at time k, the input for training the deep learning network is X _k =[x _k-n+1 ,… ,x _k-1 ,x _k ], the data length is recorded as N _data , then the obtained sample size is:

N _samples =N _data -n+1

Reorganize the data into a three-dimensional tensor form: [N _samples ,n,N _features ], where N _features represent the number of features, and in this embodiment the number of features is 3, which are current, terminal voltage and dynamic stress respectively.

In this embodiment, the data division matrix used for training the state of charge estimation model and for online testing is shown in Table 1:

Table 1 Dataset partition matrix

Step 22) Build the deep learning network structure of the battery state of charge estimation model, and train the battery state of charge estimation model offline based on the reference state of charge and the sample set.

In this embodiment, the deep learning network includes a sequence input layer, two LSTM hidden layers with 0.05 Dropout, a fully connected layer and a regression layer. Among them, each layer of LSTM selects 100 hidden units. The sequence input layer is responsible for inputting sequence data into the constructed network, the fully connected layer multiplies the output of the LSTM by the weight matrix and adds bias, and the regression layer is used to perform the regression task.

表示更新过程的候选状态，W_x表示权重，b_x是偏置，σ(·)是sigmoid激活函数，tanh(·)是双曲正切激活函数，关键激活操作如下：The hidden layer consists of multiple repeating LSTM units, and each LSTM unit transfers two states to the next unit, the unit state (c _t ) and the hidden state (h _t ). _xt and _yt are the corresponding model input and output sequences, respectively. The cell state contains the information learned from the previous time step, the hidden state is also called the output state. At each time step, the LSTM layer adds valid information to or discards invalid information from previous cell states, including the forgetting, updating, and output processes of cells and hidden states. These operations are controlled by three different gates, including the input gate _it , the forget gate _ft and the output gate o _t . The forgetting process determines which information should be discarded or kept using the forget gate. Update the level of process control cell state updates and add information to cell state using input gates. Finally, the output gate and cell state are used to decide the next hidden state.

Represents the candidate state of the update process, W _x represents the weight, b _x is the bias, σ(·) is the sigmoid activation function, tanh(·) is the hyperbolic tangent activation function, and the key activation operations are as follows:

The training data used by the LSTM model is the reference state of charge and the three-dimensional tensor sample set. The label at time k is SOC _k , and the input is X _k .

LSTM model training uses the Adam optimizer, the initial learning rate is set to 0.001, and every 100 epochs the learning rate decreases in segments, and the decrease factor is 0.2; the gradient decay factor β ₁ is 0.9, and the square gradient decay factor β ₂ is 0.999. Due to the randomness of the training process of the deep learning model, each model is trained three times, and the final SOC estimation result is the average of the three times.

Step 3) Online battery state of charge estimation

Step 31) Obtain battery current, terminal voltage and dynamic stress signals online during the actual charging and discharging process of the battery.

In vehicle applications, the battery management system collects the current, terminal voltage and dynamic stress of the battery in real time as the model input. It should be noted that the battery current, terminal voltage and dynamic stress data measured online also need to be reorganized into a three-dimensional tensor form according to step 21).

Step 32) Perform online estimation of the state of charge of the power battery according to the battery state of charge estimation model that has been trained:

SOC ^* ＝LSTM(x ^* )

Among them, LSTM is a trained battery state of charge estimation model, and x ^* is an input vector composed of battery current, terminal voltage and dynamic stress obtained online.

Figure 3 shows the state of charge estimation results of Cy14. It can be seen that in the three scenarios, the battery state of charge estimation method based on mechanical signals and deep learning can estimate the state of charge of the power battery very well. As shown in Fig. 4, the overall RMSE and MAE of state-of-charge estimation for each cycle are 1.88% and 1.35%, respectively, and the maximum value of RMSE in these cycles is lower than 5%, and the maximum value of MAE is lower than 4%, reflecting good Estimated accuracy. As shown in Table 1, working condition 4 in scenarios 1 and 2 did not participate in training in this embodiment, but small RMSE and MAE were obtained during online estimation, which shows that the method is adaptable to different working conditions.

In addition, due to the addition of stress data, compared with only current and voltage information, the estimation accuracy of the battery state of charge is increased by 0.24%. Therefore, it is less sensitive to the length of the sliding window and can obtain good results when the data length is limited. Estimate performance.

In step 3), after adding 0.5% measurement noise to the three measurement values, the state of charge estimation error has only a relative increase of 5.09% compared with the noise-free data, reflecting the good robustness of the present invention.

In summary, an embodiment of the present invention is feasible, and the error between the estimated result and the actual state of charge data is small, adaptable to different working conditions, supports the estimation of small data length, and has good robustness to measurement noise sex.

If the above functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes. .

The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall be within the scope of protection defined in the claims.

Claims (10)

1. The soft package lithium ion battery state of charge estimation method based on dynamic stress and deep learning is characterized by comprising the following steps:

step 1) establishing a battery charge-discharge data set containing dynamic stress signals, wherein the battery charge-discharge data set comprises current, terminal voltage, dynamic stress data and a reference state of charge;

step 2) off-line training of a soft package lithium ion battery state of charge estimation model based on dynamic stress and deep learning;

step 21), recombining time series characteristics of current, terminal voltage and dynamic stress in a battery charge and discharge data set into a three-dimensional tensor sample set according to an input form required by a deep learning network;

step 22), constructing a deep learning network structure of a battery state-of-charge estimation model, and offline training the battery state-of-charge estimation model based on a reference state-of-charge and a sample set;

step 3) online battery state of charge estimation:

step 31) acquiring battery current, terminal voltage and dynamic stress signals on line in the actual charging and discharging process of the battery;

step 32) carrying out on-line estimation on the state of charge of the power battery according to the trained battery state of charge estimation model.

2. The method for estimating the state of charge of the soft package lithium ion battery based on dynamic stress and deep learning according to claim 1, wherein the step 1) comprises the following steps:

step 11), constructing a soft package battery stress measurement experimental device by adopting a constraint fixture and a strain sensor;

step 12) designing a battery charge and discharge experiment according to the recommended environmental temperature and charge and discharge current of the battery, testing according to set working conditions, recording the current, terminal voltage and dynamic stress of the battery in the testing process, calculating the reference state of charge of the battery, and constructing a battery charge and discharge data set.

3. The method for estimating the state of charge of the soft-pack lithium ion battery based on dynamic stress and deep learning according to claim 2, wherein in the step 11), plane stress measurement is performed by a constraint fixture and a stress sensor in a laboratory environment, and the stress sensor is arranged between two soft-pack batteries in vehicle-mounted application to measure plane stress.

4. The method for estimating the state of charge of a soft package lithium ion battery based on dynamic stress and deep learning as claimed in claim 2, wherein in said step 12), the dynamic stress S recorded during the test is _d Is the stress related to the running state of the battery, which is the total stress S _t Subtracting the initial static stress S _s The following values:

S _d ＝S _t -S _s

wherein the total stress is an instantaneous stress measured directly on the battery surface by the pressure sensor during charge/discharge, and the static stress is a stress obtained after the battery reaches an equilibrium state for a sufficient rest time;

the battery reference state of charge is calculated by definition:

wherein SOC is _k And I _k The battery state of charge and current at time k, respectively, Δt being the sampling time, C _m Is the battery capacity.

5. The method for estimating the state of charge of the soft-pack lithium ion battery based on dynamic stress and deep learning according to claim 2, wherein in the step 12), the set working conditions include: long-term charge-discharge working condition, short-term charge-discharge working condition, pulse charge-discharge working condition and dynamic driving working condition.

6. The method for estimating a state of charge of a soft package lithium ion battery based on dynamic stress and deep learning as defined in claim 1, wherein in the step 21), the battery charge and discharge data at the k moment is x _k ＝[I _k ，V _k ，F _k ]Wherein I _k Is the current at time k, V _k Is the terminal voltage at time k, F _k Is the dynamic stress at time k;

a sliding window with the length of n is adopted to obtain the input X for training the deep learning network _k ＝[x _k-n+1 ，...，x _k-1 ，x _k ]The data length is recorded as N _data The sample size obtained is:

N _samples ＝N _daia -n+1

reorganizing the data into a three-dimensional tensor form: [ N ] _samples ，n，N _features ]Wherein N is _features Representing the number of features.

7. The method for estimating the state of charge of the soft package lithium ion battery based on dynamic stress and deep learning according to claim 1, wherein in the step 22), the deep learning network comprises an input layer, two LSTM hidden layers, a full connection layer and an output layer, and the LSTM network activation process is as follows:

wherein f _t Forgetting door, i _t Is an input door o _t Is an output door which is provided with a plurality of output doors,

is a candidate state for the update process, W _x Is the weight, b _x Is the bias, σ (·) is the sigmoid activation function, and tanh (·) is the hyperbolic tangent activation function.

8. The method for estimating the state of charge of the soft package lithium ion battery based on dynamic stress and deep learning according to claim 1, wherein in the step 32), the method for estimating the state of charge of the battery is as follows:

SOC ^* ＝LSTM(x ^* )

wherein LSTM is a trained battery state of charge estimation model, x ^* The input vector is composed of the battery current, terminal voltage and dynamic stress obtained on line.

9. A plug and play function test device for a charging pile, comprising a memory, a processor and a program stored in the memory, wherein the processor implements the method of any one of claims 1-8 when executing the program.

10. A storage medium having a program stored thereon, wherein the program, when executed, implements the method of any of claims 1-8.

Images