Slow fault diagnosis method for lithium ion battery energy storage system

Abstract

The invention discloses a lithium ion battery energy storage system slowly-varying fault diagnosis method, which comprises the following steps: establishing an equivalent circuit model of a lithium ion battery by adopting a second-order Thevenin model to obtain a system state equation and an observation equation; based on the system state equation and the observation equation, establishing a battery pack slowly-varying fault equivalent model through fault simulation; generating a fault sample by using the generative adversarial network; establishing a slowly varying fault diagnosis model through the light GBM, and training the slowly varying fault diagnosis model based on the light GBM by using the fault sample; and identifying the slow change fault of the lithium ion battery energy storage system through the trained slow change fault diagnosis model. According to the method, the equivalent circuit model of the lithium ion battery is established, the fault sample is generated by using the GAN, and the slowly-varying fault diagnosis model is established according to the lightGBM to identify the slowly-varying fault, so that the method has certain feasibility and accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of lithium ion fault diagnosis, and in particular relates to a method for diagnosing slow-changing faults of a lithium ion battery energy storage system. Background Art

[0002] Renewable energy is highly random and volatile. In order to overcome its impact on the power system, a certain proportion of energy storage systems are often configured. In recent years, electrochemical energy storage has become a popular energy storage solution due to its advantages such as high cycle efficiency, fast response speed and flexible configuration. Among electrochemical energy storage systems, lithium battery energy storage systems have the advantages of high energy density, long cycle life and good stability, and have received more and more attention.

[0003] Lithium batteries are usually connected in series or parallel to form a battery system. If the connecting parts between the batteries become loose or broken due to mechanical stress, material aging, etc., the contact resistance will increase significantly, seriously affecting the performance and life of the battery pack. At the same time, during the manufacturing process of lithium batteries, burrs on the surface of the electrode will puncture the diaphragm, causing a short circuit between the positive and negative electrodes. If timely measures are not taken, it may even lead to thermal runaway. Therefore, it is urgent to diagnose battery failures in real time and accurately in the early stage of battery failure, that is, during the slow-changing failure period, to ensure the safe and reliable operation of the lithium battery system.

[0004] At present, the battery system fault diagnosis methods are mainly divided into two types: offline and online identification. Offline diagnosis methods use electronic equipment, such as electron microscopes, etc. [8], to monitor the inside of the battery system to detect abnormal conditions. However, the cost of offline diagnosis equipment is high, and the diagnosis process cannot be repeated. Online fault diagnosis usually adopts a data-driven method. Yao et al. proposed a battery fault diagnosis method based on support vector machine (SVM) in their research. Through real-time monitoring and analysis of battery performance parameters, battery fault diagnosis was achieved. Chen et al. used deep learning algorithms to identify and classify battery faults and achieved a high accuracy rate. Sun et al. used transfer learning to optimize residual neural networks and used the trained model to detect short-circuit faults in batteries.

[0005] Data-driven methods require a large number of fault samples to train the model. However, in reality, the battery system is in a fault-free state most of the time, and fault samples are difficult to obtain. Therefore, there will be a small number of fault samples and an imbalance in the number of normal state samples, making it impossible to train an ideal machine learning model. In addition, most of the existing research can only identify single, serious faults, and there are few studies on the early detection of multiple, minor faults in battery systems. Summary of the invention

[0006] Purpose of the invention: In order to overcome the shortcomings of the prior art, a slow-changing fault diagnosis method for a lithium-ion battery energy storage system is provided. The equivalent circuit model of the lithium-ion battery is established using the second-order Thevenin model, and fault samples are generated using GAN. A slow-changing fault diagnosis model is established based on lightGBM to identify slow-changing faults, which has certain feasibility and accuracy.

[0007] Technical solution: To achieve the above purpose, the present invention provides a method for diagnosing slow-changing faults in a lithium-ion battery energy storage system, comprising the following steps:

[0008] S1: Use the second-order Thevenin model to establish the equivalent circuit model of the lithium-ion battery and obtain the system state equation and observation equation;

[0009] S2: Based on the system state equation and observation equation, an equivalent model of battery pack slow-changing fault is established through fault simulation;

[0010] S3: Based on the equivalent model of battery pack slow-changing faults, Generative Adversarial Networks (GAN) is used to solve the problem of small and unbalanced number of fault samples and generate fault samples;

[0011] S4: Establish a slow-varying fault diagnosis model through lightGBM, and use fault samples to train the slow-varying fault diagnosis model based on lightGBM;

[0012] S5: Identify slow-changing faults of lithium-ion battery energy storage systems through the trained slow-changing fault diagnosis model.

[0013] Furthermore, the step S1 is specifically as follows:

[0014] According to Kirchhoff's law, we can get:

[0015]

[0016] Calculate the battery state of charge SOC:

[0017]

[0018] Among them, SOC0 represents the initial state of charge of the battery, η represents the charge and discharge efficiency, C n Indicates battery capacity;

[0019] Discretize equations (1) and (2), and select SOC, U1, and U2 as the state variables of the system to obtain the system state equation and observation equation.

[0020] Furthermore, the system state equation and observation equation are expressed as follows:

[0021]

[0022] U out (k) = U ocv (SOC(k))-R0I(k)-U1(k)-U2(k)+v(k) (4) Wherein, Δt represents the sampling time interval and k represents the discrete time.

[0023] Furthermore, the step S2 is specifically as follows:

[0024] According to the Hybrid Pulse Power Characterization (HPPC) experiment, the battery was tested at a constant temperature and constant current of 25°C, and a group of U ocv The fitting expression can be obtained

[0025]

[0026] Among them, β9, β8, … β0 represent the coefficients to be fitted of the OCV-SOC curve.

[0027] Furthermore, the construction of the generative adversarial network in step S3 includes: the input layer of the generator and the discriminator contains 150 neurons, the number of hidden layers is 5, each hidden layer contains 21 neurons, and the relu function is used as the activation function; the output layer of the generator contains 150 neurons, and the sigmoid function is used as the activation function; the output layer of the discriminator contains 1 neuron, and the sigmoid function is also used as the activation function; the random noise adopts a uniform distribution.

[0028] Furthermore, the generation of the fault sample in step S3 includes:

[0029] Suppose the sample data is x, and its distribution is P data ; Let the noise be z, and its prior distribution be P z ; The sample generated by the generator is G(z), and the mapping space is G(z,θ g ); the output of the discriminator is the probability of judging that its input data is real data, which is D(x), and its mapping space is D(x,θ d );The objective function of GAN can be obtained

[0030]

[0031] Furthermore, in step S3, in order to improve the gradient during neural network training, the objective function of GAN is improved to obtain an improved objective function:

[0032]

[0033] Furthermore, in step S3, the improved objective function is optimized by alternating iterations: first, the generator is kept unchanged, iterative training is performed K times, and the loss function of the discriminator is minimized; then the discriminator is kept unchanged, iterative training is performed 1 time, and the loss function of the generator is minimized, and the above process is repeated until the stopping condition is met.

[0034] Furthermore, the specific steps of alternately iteratively optimizing the improved objective function in step S3 are:

[0035] ①Let k = 1, l = 1;

[0036] ② Randomly select M sample data, obeying the distribution x~P data , and then extract M noise data, obeying the distribution z~P z ;

[0037] ③ Input the extracted 2M samples into the discriminator D, calculate the discriminator's objective function according to formula (7), and use the gradient descent algorithm to train the discriminator parameters θ d ;

[0038] ④ Set k = k + 1. If k > K, jump to ⑤ and set k = 1 again. Otherwise, go back to ②.

[0039] ⑤ Extract M noise data, obeying the distribution z~P z ;

[0040] ⑥ Input the extracted M data into the generator G, calculate the objective function of the generator according to formula (8), and also use the gradient descent algorithm to train the generator parameter θ g ;

[0041] ⑦l=l+1, if l>loop GAN Then exit the loop, otherwise return to ②.

[0042] Furthermore, the training method of the gradient descent algorithm in steps ③ and ⑥ is: the main idea of ​​the gradient boosting decision tree (GBDT) is to generate multiple weak learners - classification and regression trees (CART) by continuously iteratively training the negative gradient of the loss function, and then obtain the final strong learner by linearly combining multiple weak learners.

[0043]

[0044] Among them, F(x) is the output value of the strong learner, f0(x) is the output value of the initial learner, and f m(x) is the output value of the mth weak learner, and M is the number of weak learners.

[0045] Beneficial effects: Compared with the prior art, the present invention can make up for the situation that the fault samples of the battery system are difficult to obtain in actual situations, and the fault sample data is much smaller than the normal sample data. In addition, most of the prior art can only identify single, serious faults, while the present invention can effectively identify multiple, minor faults in the early stage of the battery system, improving the diagnostic effect of slow-changing faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the diagnostic method of the present invention;

[0047] Figure 2 This is the second-order RC circuit model diagram of a lithium battery;

[0048] Figure 3 It is the equivalent circuit diagram of slow-changing fault;

[0049] Figure 4 It is a schematic diagram of the slow-changing fault simulation experiment;

[0050] Figure 5 This is the GAN flow chart;

[0051] Figure 6 It is the SOC-OCV curve fitting result diagram;

[0052] Figure 7 It is the GAN loss function graph. DETAILED DESCRIPTION

[0053] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0054] like Figure 1 As shown, the present invention provides a method for diagnosing slow-changing faults of a lithium-ion battery energy storage system, comprising the following steps:

[0055] S1: Use the second-order Thevenin model to establish the equivalent circuit model of the lithium-ion battery and obtain the system state equation and observation equation;

[0056] According to Kirchhoff's law, we can get:

[0057]

[0058] Calculate the battery state of charge SOC:

[0059]

[0060] Among them, SOC0 represents the initial state of charge of the battery, η represents the charge and discharge efficiency, C n Indicates battery capacity;

[0061] Discretize equations (1) and (2), and select SOC, U1, and U2 as the state variables of the system to obtain the system state equation and observation equation;

[0062] The system state equation and observation equation are expressed as follows:

[0063]

[0064] U out (k) = U ocv (SOC(k))-R0I(k)-U1(k)-U2(k)+v(k) (4)

[0065] Wherein, Δt represents the sampling time interval, which is set to 1s, and k represents the discrete time.

[0066] Since complex chemical reactions occur inside lithium-ion batteries when they are charged and discharged, the present invention adopts a second-order Thevenin model in order to balance accuracy and complexity.

[0067] In this embodiment, the Figure 2 Equivalent circuit model of a lithium-ion battery is shown.

[0068] S2: Based on the system state equation and observation equation, an equivalent model of battery pack slow-changing fault is established through fault simulation;

[0069] The slow-changing faults of lithium battery packs mainly include micro-short circuit faults inside the battery and connection faults between batteries. In order to better diagnose the slow-changing faults, the characteristics of the two faults are analyzed. This paper establishes equivalent circuit models of the two faults, such as Figure 3 As shown, the left side is the equivalent circuit diagram of the battery pack micro-short circuit fault, and the right side is the equivalent circuit diagram of the battery pack connection fault.

[0070] Fault simulation experiment Figure 4 As shown, the equivalent model of the battery pack slow-changing fault can be quickly established, which is easy to simulate and realize, has high repeatability and controllability, and makes up for the shortage of actual slow-changing fault samples.

[0071] S3: Based on the equivalent model of battery pack slow-changing faults, Generative Adversarial Networks (GAN) is used to solve the problem of small and unbalanced number of fault samples and generate fault samples;

[0072] According to the Hybrid Pulse Power Characterization (HPPC) experiment, the battery was tested at a constant temperature and constant current of 25°C, and a group of U ocv The fitting expression can be obtained

[0073]

[0074] Among them, β9, β8, … β0 represent the coefficients to be fitted of the OCV-SOC curve.

[0075] The construction of the generative adversarial network includes: the input layer of the generator and the discriminator contains 150 neurons, the number of hidden layers is 5, each hidden layer contains 21 neurons, and the relu function is used as the activation function; the output layer of the generator contains 150 neurons, and the sigmoid function is used as the activation function; the output layer of the discriminator contains 1 neuron, and the sigmoid function is also used as the activation function; the random noise adopts a uniform distribution.

[0076] like Figure 5 As shown, the generation of fault samples includes:

[0077] Suppose the sample data is x, and its distribution is P data ; Let the noise be z, and its prior distribution be P z ; The sample generated by the generator is G(z), and the mapping space is G(z,θ g ); the output of the discriminator is the probability of judging that its input data is real data, which is D(x), and its mapping space is D(x,θ d );The objective function of GAN can be obtained

[0078]

[0079] In order to improve the gradient during neural network training, the objective function of GAN is improved, and the improved objective function is obtained

[0080]

[0081] The improved objective function is optimized through alternating iterations: first keep the generator unchanged, iterate and train K times to minimize the loss function of the discriminator; then keep the discriminator unchanged, iterate and train 1 time to minimize the loss function of the generator, and repeat the above process until the stopping condition is met.

[0082] The specific steps of alternating iterative optimization to improve the objective function are:

[0083] ①Let k = 1, l = 1;

[0084] ② Randomly select M sample data, obeying the distribution x~Pdata , and then extract M noise data, obeying the distribution z~P z ;

[0085] ③ Input the extracted 2M samples into the discriminator D, calculate the discriminator's objective function according to formula (7), and use the gradient descent algorithm to train the discriminator parameters θ d ;

[0086] ④ Set k = k + 1. If k > K, jump to ⑤ and set k = 1 again. Otherwise, go back to ②.

[0087] ⑤ Extract M noise data, obeying the distribution z~P z ;

[0088] ⑥ Input the extracted M data into the generator G, calculate the objective function of the generator according to formula (8), and also use the gradient descent algorithm to train the generator parameter θ g ;

[0089] ⑦l=l+1, if l>loop GAN Then exit the loop, otherwise return to ②.

[0090] The training method using the gradient descent algorithm is: The main idea of ​​the Gradient Boosting Decision Tree (GBDT) is to generate multiple weak learners - Classification and Regression Trees (CART) by continuously iteratively training the negative gradient of the loss function, and then obtain the final strong learner by linearly combining multiple weak learners.

[0091]

[0092] Among them, F(x) is the output value of the strong learner, f0(x) is the output value of the initial learner, and f m (x) is the output value of the mth weak learner, and M is the number of weak learners.

[0093] S4: Establish a slow-varying fault diagnosis model through lightGBM, and use fault samples to train the slow-varying fault diagnosis model based on lightGBM;

[0094] S5: Identify two slow-changing faults of lithium-ion battery energy storage system, micro-short circuit fault and connection fault, through the trained slow-changing fault diagnosis model.

[0095] In order to verify the effectiveness and effect of the method of the present invention, the following experiments and analyses were performed in this embodiment, as follows:

[0096] This example analyzes the equivalent circuit model of lithium-ion batteries. In order to balance accuracy and complexity, a second-order RC circuit model of lithium-ion batteries is constructed, SOC, U1, and U2 are selected as the state variables of the system, and discretized to establish the system state equation and observation equation, where the battery SOC value is related to the open circuit voltage U ocv There is a nonlinear relationship, and the SOC-OCV curve fitting results are as follows Figure 6 shown.

[0097] In the slow-changing fault simulation experiment, the battery cell uses a 2.8Ah battery with unknown parameters. The battery pack is composed of 8 identical 2.8Ah batteries connected in series. Battery No. 1 is simulated as a faulty battery cell, and the other 7 batteries are normal batteries. The micro-short circuit fault resistance value range is 50-500Ω, with an interval of 1Ω, a total of 451 resistance values, and the connection fault resistance value is 0.05-1mΩ, with an interval of 0.001mΩ, a total of 996 resistance values. For each fault experiment, the battery pack is first discharged to the discharge cut-off voltage before the experiment is carried out. The experiment was carried out in an environment of 25℃, first with a constant current of 0.5C charging to the charging cut-off voltage, and then with a constant current of 0.5C discharging to the discharge cut-off voltage. After the experiment, a total of 451 micro-short circuit battery data, 996 connection fault data, and 10129 normal battery data can be obtained. In this embodiment, the sampling period of the original voltage curve is 1 second. The complete charge and discharge process has about 15,000 data points. If the data is sampled every 100 seconds, there are 150 input features in total.

[0098] To simulate the situation in which the battery fault data is less than the normal data in actual conditions, this embodiment randomly selects 200 pieces of 10129 pieces of normal data as training samples, and randomly selects 100 pieces of each type of fault state as training samples.

[0099] Construct GAN: The input layer of the generator and the discriminator contains 150 neurons, the number of hidden layers is 5, each hidden layer contains 21 neurons, and the relu function is used as the activation function. The output layer of the generator contains 150 neurons, and the sigmoid function is used as the activation. The output layer of the discriminator contains 1 neuron, and the sigmoid function is also used as the activation. The random noise is uniformly distributed. This embodiment uses the RMSProp algorithm to optimize the gradient descent method to enhance the stability of model training, and the learning rate is 0.001. The number of samples randomly taken each time is M=16, and the number of iterations is loop. GAN =10 4 , take K=2.

[0100] Based on the 100 fault training samples, GAN is used to generate 100 generated samples for each type of fault, for a total of 200 training samples, so that the training samples of each type of state are balanced. Taking the generation of micro short circuit fault samples as an example, the loss function values ​​of D and G are recorded once every 10 iterations of GAN and plotted on Figure 7 middle.

[0101] from Figure 7 It can be seen that the loss functions of the generator G and the discriminator D are also constantly fluctuating, indicating that the two are competing. As the iteration proceeds, the loss function tends to stabilize, indicating that the competition has reached a balanced state.

[0102] The lightGBM-based slow-changing fault diagnosis model is trained using balanced training samples. Then, 200 data are randomly selected from all the data in each state as test samples to verify the effectiveness of the fault diagnosis method proposed in this paper. In the normal state, 194 groups of samples were successfully diagnosed; in the micro-short circuit fault, 186 groups of samples were successfully diagnosed; in the connection fault, 181 groups of samples were successfully diagnosed, and the overall recognition accuracy was 93.5%.

[0103] From the results, it can be seen that the battery pack slow-changing fault diagnosis based on GAN-lightGBM proposed in the present invention is effective and has a good diagnostic effect.

Claims

1. A method for diagnosing slow-changing faults in a lithium-ion battery energy storage system, characterized in that: The steps include: S1: Use the second-order Thevenin model to establish the equivalent circuit model of the lithium-ion battery and obtain the system state equation and observation equation; S2: Based on the system state equation and observation equation, an equivalent model of battery pack slow-changing fault is established through fault simulation; S3: Based on the equivalent model of battery pack slow-changing faults, fault samples are generated using a generative adversarial network; S4: Establish a slow-varying fault diagnosis model through lightGBM, and use fault samples to train the slow-varying fault diagnosis model based on lightGBM; S5: Identify slow-changing faults of lithium-ion battery energy storage systems through the trained slow-changing fault diagnosis model.

2. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 1, characterized in that: The step S1 is specifically as follows: According to Kirchhoff's law, we can get: Calculate the battery state of charge SOC: Among them, SOC0 represents the initial state of charge of the battery, η represents the charge and discharge efficiency, C n Indicates the battery capacity; Discretize equations (1) and (2), and select SOC, U1, and U2 as the state variables of the system to obtain the system state equation and observation equation.

3. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 2, characterized in that: The system state equation and observation equation are expressed as follows: U out (k)=U ocv (SOC(k))-R0I(k)-U1(k)-U2(k)+v(k) (4) Where Δt represents the sampling time interval and k represents the discrete time.

4. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 3, characterized in that: The step S2 is specifically as follows: According to the hybrid pulse power test experiment, the battery was tested under constant temperature and constant current conditions, and a set of U ocv The fitting expression can be obtained Among them, β9, β8, ... β0 represent the coefficients to be fitted of the OCV-SOC curve.

5. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 3, characterized in that: The construction of the generative adversarial network in step S3 includes: the input layer of the generator and the discriminator contains 150 neurons, the number of hidden layers is 5, each hidden layer contains 21 neurons, and the relu function is used as the activation function; the output layer of the generator contains 150 neurons, and the sigmoid function is used as the activation function; the output layer of the discriminator contains 1 neuron, and the sigmoid function is also used as the activation function; the random noise adopts a uniform distribution.

6. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 5, characterized in that: The generation of the fault sample in step S3 includes: Suppose the sample data is x, and its distribution is P data ; Let the noise be z, and its prior distribution be P z ; The sample generated by the generator is G(z), and the mapping space is G(z,θ g ); the output of the discriminator is the probability of judging that its input data is real data, which is D(x), and its mapping space is D(x,θ d );The objective function of GAN can be obtained 7. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 6, characterized in that: In step S3, the objective function of GAN is improved to obtain an improved objective function:

8. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 7, characterized in that: In the step S3, the improved objective function is optimized by alternating iterations: first, the generator is kept unchanged, iterative training is performed K times, and the loss function of the discriminator is minimized; then, the discriminator is kept unchanged, iterative training is performed 1 time, and the loss function of the generator is minimized, and the above process is repeated until the stopping condition is met.

9. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 8, characterized in that: The specific steps of alternately iterating and optimizing the improved objective function in step S3 are: ①Let k = 1, l = 1; ② Randomly select M sample data, obeying the distribution x~P data , and then extract M noise data, obeying the distribution z~P z ; ③ Input the extracted 2M samples into the discriminator D, calculate the discriminator objective function according to formula (7), and use the gradient descent algorithm to train the discriminator parameters θ d ; ④ Set k = k + 1. If k > K, jump to ⑤ and set k = 1 again. Otherwise, go back to ②. ⑤ Extract M noise data, obeying the distribution z~P z ; ⑥ Input the extracted M data into the generator G, calculate the objective function of the generator according to formula (8), and also use the gradient descent algorithm to train the generator parameter θ g ; ⑦l=l+1, if l>loop GAN Then exit the loop, otherwise return to ②.

10. A method for diagnosing slow-change faults in a lithium-ion battery energy storage system according to claim 9, characterized in that: The training method of the gradient descent algorithm in steps ③ and ⑥ is: by continuously iteratively training the negative gradient of the loss function, multiple weak learners - classification regression trees are generated, and then multiple weak learners are linearly combined to obtain the final strong learner. Among them, F(x) is the output value of the strong learner, f0(x) is the output value of the initial learner, and f m (x) is the output value of the mth weak learner, and M is the number of weak learners.

Images