CNN-BiGRU neural network-based all-vanadium redox flow battery fault diagnosis method

Through the method based on the CNN-BiGRU neural network, a vanadium battery fault diagnosis model is constructed, which solves the problem of time-consuming and experimental complexity of vanadium battery fault data acquisition, and realizes real-time fault diagnosis and high-precision judgment of vanadium battery.

CN118938015APending Publication Date: 2024-11-12WUHAN UNIV OF TECH

Patent Information

Application Number
CN202411049526.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The prior art is difficult to effectively diagnose vanadium battery failure, especially because vanadium battery failure data collection is time-consuming and cumbersome and fluid leakage is prone to occur during the experiment, making it difficult to build an applicable fault diagnosis model.

Method used

Using a method based on CNN-BiGRU neural network, battery data is collected through multi-condition charging and discharging experiments, offline voltage prediction model is constructed, and the final voltage prediction model is obtained through online learning to judge battery failure in real time, avoiding the cumbersome collection of traditional fault data. process.

Benefits of technology

Real-time fault diagnosis of vanadium batteries is realized, diagnostic accuracy is improved, false alarm rate is reduced, no additional sensor support is required, and time-consuming of complex fault experiments is avoided.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118938015A_ABST
    Figure CN118938015A_ABST
Patent Text Reader

Abstract

The invention discloses an all-vanadium redox flow battery fault diagnosis method based on a CNN-BiGRU neural network. The method comprises the following steps that experiment working conditions are designed, sample data under charging and discharging experiments of all the working conditions are collected, and each piece of sample data is composed of battery voltage, current, flow and SOC data; preprocessing the data; building a CNN-BiGRU neural network model, training the CNN-BiGRU neural network model to obtain an off-line voltage prediction model, evaluating the CNN-BiGRU neural network model, and optimizing partial hyper-parameters of the model; performing online learning on the off-line voltage model to obtain a final voltage prediction model; obtaining a predicted voltage value from the final voltage prediction model, and subtracting the operation voltage value from the predicted voltage value to obtain a voltage residual error; and comparing the voltage residual error with a fault threshold, and judging whether the new battery has a fault or not. The fault diagnosis model suitable for the vanadium battery is constructed, the fault diagnosis precision is high, and the false alarm rate is low.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of all-vanadium liquid flow batteries, and in particular to an all-vanadium liquid flow battery fault diagnosis method based on a CNN-BiGRU neural network. Background Art

[0002] Vanadium batteries stand out among many energy storage technologies for their long cycle life, independence of energy and power, and environmental friendliness. Vanadium batteries can provide stable, reliable, and long-life energy storage solutions, and are an indispensable part of future energy transformation and smart grid development.

[0003] The all-vanadium liquid flow battery is a complex system coupled with multiple physical fields. During its operation, it may experience faults such as proton exchange membrane damage, pump failure, and leakage. Among these faults, pump failure is the most serious, which will cause the battery voltage to change rapidly and cause irreversible damage to the battery stack. Other faults will also reduce the operating efficiency of the all-vanadium liquid flow battery and increase operating costs. Therefore, a reasonable and effective fault diagnosis strategy is crucial for all-vanadium liquid flow batteries.

[0004] In recent years, the rapid development of artificial intelligence has promoted the application of neural networks in the field of battery fault diagnosis. Previous literature mainly focused on the fault diagnosis of lithium batteries and fuel cells, such as: Chinese patent CN117388702A discloses a fuel cell fault pre-diagnosis model, which judges the type of subsequent possible faults of the battery by inputting the operating status data of the fuel cell; Chinese patent CN110308397A discloses a model for diagnosing lithium batteries based on Fourier transform and convolutional neural network.

[0005] However, the method of collecting fault data and then building a fault diagnosis model is not suitable for vanadium batteries, because for vanadium batteries, there are the following problems in collecting fault data:

[0006] First, for large battery stacks, simulating battery failures often requires removing wires and installing the battery stack, which is very time-consuming and cumbersome.

[0007] Second, for single batteries, leakage problems are prone to occur during the experiment, and under certain faults, the single battery may not be able to continue to operate. Summary of the invention

[0008] In view of the shortcomings of the prior art, the present invention proposes a vanadium liquid flow battery fault diagnosis method based on a CNN-BiGRU neural network to achieve real-time fault diagnosis during the operation of the vanadium battery.

[0009] To achieve the above object, the present invention designs a vanadium liquid flow battery fault diagnosis method based on CNN-BiGRU neural network, which is particularly characterized by comprising the following steps:

[0010] S1) performing multi-condition charge and discharge experiments on the all-vanadium liquid flow battery, and collecting sample data under each condition of the charge and discharge experiment according to a set sampling frequency, each of the sample data is composed of battery voltage, current, flow rate and SOC data;

[0011] S2) dividing the collected sample data into training set data and test set data, respectively, using the current, flow and SOC data in the training set data and the test set data as the input features of the neural network, and using the voltage in the training set data and the test set data as the output features of the neural network; and performing normalization operations on each feature in the training set data and the test set data one by one, and then performing slicing processing;

[0012] S3) building a CNN-BiGRU neural network model, training the CNN-BiGRU neural network model using the training set data to obtain an offline voltage prediction model, evaluating the CNN-BiGRU neural network model using the test set data, and optimizing some hyperparameters of the model;

[0013] S4) deploying the offline voltage prediction model before the new all-vanadium liquid flow battery, collecting some new battery operation data, and performing online learning on the offline voltage model to obtain the final voltage prediction model;

[0014] S5) Input the real-time flow rate, current and SOC of the new all-vanadium liquid flow battery into the final voltage prediction model to obtain the predicted voltage value, subtract the predicted voltage value from the operating voltage value of the new all-vanadium liquid flow battery actually collected by the sensor, and use the absolute value of the difference as the voltage residual; if the voltage residual is greater than the set fault threshold, it is determined that the new battery has a fault; otherwise, the new battery is normal.

[0015] Further, in S1), the operating conditions of the charge and discharge experiment include constant flow and constant current conditions, constant flow and variable current conditions, and variable flow and variable current conditions, and the sampling frequency is set to 1HZ.

[0016] Furthermore, in S1), the battery SOC is calculated by the ampere-hour integration method

[0017]

[0018] ΔCapacity=∫Idt

[0019] In the formula,

[0020] SOC is the state of charge of the battery, that is, the remaining capacity of the battery.

[0021] SOC0 is the initial state of charge of the battery.

[0022] ΔCapacity is the change in battery capacity,

[0023] Capacity norminal is the nominal capacity of the battery,

[0024] I is the current in t seconds,

[0025] t is the sampling frequency.

[0026] Furthermore, in S1), a small current charge and discharge experiment is performed on the battery to obtain the battery's OCV-SOC curve and the battery's nominal capacity Capacity norminal , the initial state of charge SOC0 of the battery is obtained from the OCV-SOC curve of the battery.

[0027] Furthermore, in S2), the features in the training set data and the test set data are normalized by the following formula:

[0028]

[0029] In the formula,

[0030] x N is the normalized eigenvalue,

[0031] x min is the minimum value of a feature in the training set / test set,

[0032] x max It is the maximum value of a feature in the training set / test set.

[0033] Furthermore, in S3), a CNN-BiGRU neural network model is built in a Pytorch environment, wherein the CNN-BiGRU neural network model consists of an input layer, a convolutional layer, two BiGRU layers, a dropout layer, and a fully connected layer.

[0034] Furthermore, in S3), the evaluation index for evaluating the CNN-BiGRU neural network model using the test set data is

[0035]

[0036]

[0037]

[0038] In the formula,

[0039] MAE is the first evaluation indicator. The smaller the value, the higher the model accuracy.

[0040] MSE is the second evaluation indicator. The smaller the value, the higher the model accuracy.

[0041] MAPE is the third evaluation indicator. The smaller the value, the higher the model accuracy.

[0042] n is the number of samples in the test set data,

[0043] is the predicted voltage obtained by the offline voltage prediction model,

[0044] V is the actual voltage in the test set data.

[0045] Furthermore, in S3), the hyperparameters of the CNN-BiGRU neural network model are optimized, including the sample length s and the batch size B in the model training. The CNN-BiGRU neural network model is trained under different values ​​of the sample length s and the batch size B to obtain the evaluation indicators of each model, and the optimal s and B are selected according to the evaluation indicators.

[0046] Furthermore, it also includes step S6), when the capacity of the new all-vanadium liquid flow battery decays by 10%, the final voltage prediction model is re-learned online to obtain the latest state of the new all-vanadium liquid flow battery after capacity decay, thereby improving the prediction accuracy of the final voltage prediction model and reducing the false alarm rate of faults.

[0047] The advantages of the present invention are:

[0048] 1. The present invention trains the CNN-BiGRU neural network model to obtain an offline voltage prediction model, and then performs online learning on the offline voltage model to obtain a final voltage prediction model. The predicted voltage value output by the final voltage prediction model is subtracted from the actual operating voltage value collected to obtain a voltage residual. Whether the battery is faulty is determined by comparing the voltage residual with a given threshold;

[0049] 2. The diagnostic method of the present invention can effectively diagnose battery faults without the need for various sensors, and has high accuracy and low false alarm rate;

[0050] 3. The diagnostic method of the present invention does not need to collect various data of battery failure, thus avoiding the complicated and time-consuming failure experiment process of vanadium batteries;

[0051] The present invention is based on the CNN-BiGRU neural network all-vanadium liquid flow battery fault diagnosis method, through the CNN-BiGRU neural network model and online learning, to build a fault diagnosis model suitable for vanadium batteries, the fault diagnosis accuracy is high, the false alarm rate is low. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the present invention;

[0053] Figure 2A comparison diagram of the predicted voltage value and the actual operating voltage value of the model in the embodiment;

[0054] Figure 3 A comparison diagram of the predicted voltage value and the actual operating voltage value when the pump fails in the embodiment;

[0055] Figure 4 Graph showing the voltage residual variation when a pump fails in the embodiment. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, the present invention provides a vanadium liquid flow battery fault diagnosis method based on a CNN-BiGRU neural network, comprising the following steps:

[0058] S1) performing multi-condition charge and discharge experiments on the all-vanadium liquid flow battery, and collecting sample data under each condition of the charge and discharge experiment according to a set sampling frequency, each of the sample data is composed of battery voltage, current, flow rate and SOC data.

[0059] The parameters of the all-vanadium liquid flow battery in this example are shown in Table 1.

[0060] Table 1 Parameters of the all-vanadium liquid flow battery experimental platform

[0061]

[0062]

[0063] Design the operating conditions of vanadium batteries, which should reflect the operating characteristics of vanadium batteries as comprehensively as possible. Collect the current, flow, SOC and voltage data of the battery under various operating conditions.

[0064] Specifically, the operating conditions of the charge and discharge experiment include a constant flow and constant current operating condition, a constant flow and variable current operating condition, and a variable flow and variable current operating condition, and the sampling frequency is set to 1 Hz.

[0065] The working conditions in this embodiment are as follows:

[0066] Single cycle constant flow and constant current condition: the battery current was set to a constant value of 1A, and 6 cycles of charge and discharge experiments were carried out. The flow rates of each cycle were set to 15mL / min, 30mL / min, 45mL / min, 60mL / min, 75mL / min, and 90mL / min respectively.

[0067] Single cycle constant flow variable current condition: the battery current changes linearly in the range of 0.6-1.8A, and 6 cycles of charge and discharge experiments are carried out. The flow rate of each cycle is set to 15mL / min, 30mL / min, 45mL / min, 60mL / min, 75mL / min, and 90mL / min respectively.

[0068] Single cycle variable flow and variable current conditions: In a single cycle, the battery current varies in the range of 0.6 to 1.8 A, and the flow rate varies in the range of 15 to 90 mL / min.

[0069] Specifically, the battery SOC is calculated by the ampere-hour integration method

[0070]

[0071] ΔCapacity=∫Idt

[0072] In the formula,

[0073] SOC is the state of charge of the battery, that is, the remaining capacity of the battery.

[0074] SOC0 is the initial state of charge of the battery.

[0075] ΔCapacity is the change in battery capacity,

[0076] Capacity norminal is the nominal capacity of the battery,

[0077] I is the current in t seconds,

[0078] t is the sampling frequency, and the sampling frequency in this embodiment is 1 second.

[0079] Perform a low current charge and discharge experiment on the battery to obtain the battery's OCV-SOC curve and the battery's nominal capacity. norminal , the initial state of charge SOC0 of the battery is obtained from the OCV-SOC curve of the battery.

[0080] The nominal capacity in this example is norminal It is 1107mAh.

[0081] S2) The collected sample data are divided into training set data and test set data, respectively, and the current, flow and SOC data in the training set data and the test set data are used as the input features of the neural network, and the voltage in the training set data and the test set data are used as the output features of the neural network; and each feature in the training set data and the test set data is normalized one by one, and then sliced.

[0082] Specifically, the features in the training set data and the test set data are normalized by the following formula:

[0083]

[0084] In the formula,

[0085] x N is the normalized eigenvalue,

[0086] x min is the minimum value of a feature in the training set / test set,

[0087] x max It is the maximum value of a feature in the training set / test set.

[0088] In this embodiment, the slice length for slicing the feature is set to t, the single sample dimension of the input feature is [t, 3], and the single sample dimension of the output feature is [t, 1].

[0089] S3) Build a CNN-BiGRU neural network model, use the training set data to train the CNN-BiGRU neural network model to obtain an offline voltage prediction model, use the test set data to evaluate the CNN-BiGRU neural network model, and optimize some hyperparameters of the model.

[0090] Specifically, a CNN-BiGRU neural network model is built in a Pytorch environment, and the CNN-BiGRU neural network model consists of an input layer, a convolutional layer, two BiGRU layers, a dropout layer, and a fully connected layer.

[0091] Specifically, the evaluation index for evaluating the CNN-BiGRU neural network model using the test set data is

[0092]

[0093]

[0094]

[0095] In the formula,

[0096] MAE is the first evaluation indicator. The smaller the value, the higher the model accuracy.

[0097] MSE is the second evaluation indicator. The smaller the value, the higher the model accuracy.

[0098] MAPE is the third evaluation indicator. The smaller the value, the higher the model accuracy.

[0099] n is the number of samples in the test set data,

[0100] is the predicted voltage obtained by the offline voltage prediction model,

[0101] V is the actual voltage in the test set data.

[0102] Specifically, the hyperparameters of the CNN-BiGRU neural network model are optimized, including the sample length s and the batch size B in model training. The CNN-BiGRU neural network model is trained under different values ​​of the sample length s and the batch size B to obtain the evaluation indicators of each model, and the optimal s and B are selected according to the evaluation indicators.

[0103] In this embodiment, the interval of s is set to [5, 10, 15, 20, 25, 30, 35, 40], and the value interval of B is [4, 8, 16, 32, 64, 128]. The model is trained under different values ​​of s and B to obtain the evaluation indicators of each model. Here, the optimal sample length s is 5 and the batch size B is 16. In addition, the failure of the all-vanadium liquid flow battery is not particularly severe, and predicting the voltage every 5 seconds for diagnosis is sufficient for vanadium batteries.

[0104] S4) Deploy the offline voltage prediction model before the new all-vanadium liquid flow battery, collect some new battery operation data, conduct online learning on the offline voltage model, and obtain the final voltage prediction model.

[0105] Follow these steps to train the model online:

[0106] A small amount of operating data of the newly deployed batteries should be used. The new data should reflect the polarization characteristics and internal resistance characteristics of the batteries as much as possible. Based on the trained offline model, the fully connected layers should be retrained, and the weight parameters of other layers should be frozen. If necessary, the fully connected layers and some BiGRU layers can be retrained. In addition, the learning rate during training should be lowered.

[0107] In this embodiment, the collected operating data of the new battery is a set of constant current charge and discharge data and a set of random pulse operating condition data of the new battery. Figure 2 As shown in the figure, it is a comparison chart between the model predicted voltage value and the actual operating voltage value. The predicted voltage value is highly consistent with the actual operating voltage value, and the model accuracy is very high.

[0108] S5) Input the real-time flow rate, current and SOC of the new all-vanadium liquid flow battery into the final voltage prediction model to obtain the predicted voltage value, subtract the predicted voltage value from the operating voltage value of the new all-vanadium liquid flow battery actually collected by the sensor, and use the absolute value of the difference as the voltage residual; if the voltage residual is greater than the set fault threshold, it is determined that the new battery has a fault; otherwise, the new battery is normal.

[0109] The voltage residual is calculated by the following formula

[0110] Res=|V act -V pre |

[0111] In the formula,

[0112] Res is the voltage residual,

[0113] V act is the actual operating voltage value,

[0114] V pre is the predicted voltage value.

[0115] By analyzing the residuals of the non-fault data, a suitable fault threshold is set as 50 mV in this embodiment.

[0116] Here, the model is tested using a pump failure as an example. At 616 seconds, the left pump of the battery is turned off to simulate a battery pump failure, and the pump is restarted at 736 seconds. The predicted voltage and the actual voltage are shown in Figure 3 As shown in the figure, when the pump fails, the battery voltage will change rapidly. This is because the electrolyte cannot flow into the battery after the pump is turned off. The battery can only rely on the small amount of electrolyte stored in the battery for charging and discharging, and the voltage change range is relatively large. However, the voltage predicted by the neural network model is only determined by the battery current, flow rate and SOC, and the battery SOC is obtained by the ampere-hour integration method, so the model's predicted voltage will not be affected by the fault. The voltage residual is as follows Figure 4 As shown in FIG. 1 , the voltage residual gradually increases from the 616th second until it reaches the set threshold of 50 mV, triggering a fault alarm.

[0117] The vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network of the present invention also includes step S6), when the capacity of the new vanadium liquid flow battery decays by 10%, the final voltage prediction model is re-learned online to obtain the latest state of the new vanadium liquid flow battery after capacity decay, thereby improving the prediction accuracy of the final voltage prediction model and reducing the fault false alarm rate.

[0118] The present invention is based on the CNN-BiGRU neural network all-vanadium liquid flow battery fault diagnosis method, through the CNN-BiGRU neural network model and online learning, to build a fault diagnosis model suitable for vanadium batteries, the fault diagnosis accuracy is high, the false alarm rate is low.

[0119] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A vanadium liquid flow battery fault diagnosis method based on CNN-BiGRU neural network, characterized in that: The steps include: S1) performing multi-condition charge and discharge experiments on the all-vanadium liquid flow battery, and collecting sample data under each condition of the charge and discharge experiment according to a set sampling frequency, each of the sample data is composed of battery voltage, current, flow rate and SOC data; S2) dividing the collected sample data into training set data and test set data, respectively, using the current, flow rate and SOC data in the training set data and the test set data as input features of the neural network, and using the voltage in the training set data and the test set data as output features of the neural network; Each feature in the training set data and the test set data is normalized one by one, and then sliced; S3) building a CNN-BiGRU neural network model, training the CNN-BiGRU neural network model using the training set data to obtain an offline voltage prediction model, evaluating the CNN-BiGRU neural network model using the test set data, and optimizing some hyperparameters of the model; S4) deploying the offline voltage prediction model before the new all-vanadium liquid flow battery, collecting some new battery operation data, and performing online learning on the offline voltage model to obtain the final voltage prediction model; S5) inputting the real-time flow rate, current and SOC of the all-vanadium liquid flow new battery into the final voltage prediction model to obtain a predicted voltage value, subtracting the predicted voltage value from the operating voltage value of the all-vanadium liquid flow new battery actually collected by the sensor, and taking the absolute value of the difference as the voltage residual; If the voltage residual is greater than the set fault threshold, the new battery is determined to be faulty; otherwise, the new battery is normal.

2. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 1, characterized in that: S1), the operating conditions of the charge and discharge experiment include constant flow and constant current conditions, constant flow and variable current conditions, and variable flow and variable current conditions, and the sampling frequency is set to 1HZ.

3. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 2 is characterized in that: In S1), the battery SOC is calculated by the ampere-hour integration method. ΔCapacity=∫Idt In the formula, SOC is the state of charge of the battery, that is, the remaining capacity of the battery. SOC0 is the initial state of charge of the battery. ΔCapacity is the change in battery capacity, Capacity norminal is the nominal capacity of the battery, I is the current in t seconds, t is the sampling frequency.

4. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 3 is characterized in that: In S1), a small current charge and discharge experiment is performed on the battery to obtain the battery's OCV-SOC curve and the battery's nominal capacity. norminal , the initial state of charge SOC0 of the battery is obtained from the OCV-SOC curve of the battery.

5. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 1, characterized in that: In S2), the features in the training set data and the test set data are normalized by the following formula: In the formula, x N is the normalized eigenvalue, x min is the minimum value of a feature in the training set / test set, x max It is the maximum value of a feature in the training set / test set.

6. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 1, characterized in that: S3), a CNN-BiGRU neural network model is built in a Pytorch environment, wherein the CNN-BiGRU neural network model consists of an input layer, a convolutional layer, two BiGRU layers, a dropout layer, and a fully connected layer.

7. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 6 is characterized in that: In S3), the evaluation index for evaluating the CNN-BiGRU neural network model using the test set data is In the formula, MAE is the first evaluation indicator. The smaller the value, the higher the model accuracy. MSE is the second evaluation indicator. The smaller the value, the higher the model accuracy. MAPE is the third evaluation indicator. The smaller the value, the higher the model accuracy. n is the number of samples in the test set data, is the predicted voltage obtained by the offline voltage prediction model, V is the actual voltage in the test set data.

8. The all-vanadium liquid flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 7, characterized in that: In S3), the hyperparameters of the CNN-BiGRU neural network model are optimized, including the sample length s and the batch size B in the model training. The CNN-BiGRU neural network model is trained under different values ​​of the sample length s and the batch size B to obtain the evaluation indicators of each model, and the optimal s and B are selected according to the evaluation indicators.

9. The all-vanadium redox flow battery fault diagnosis method based on the CNN-BiGRU neural network according to claim 1, characterized in that: The method also includes step S6), in which the final voltage prediction model is re-learned online every time the capacity of the new all-vanadium liquid flow battery decays by 10% to obtain the latest state of the new all-vanadium liquid flow battery after capacity decay, thereby improving the prediction accuracy of the final voltage prediction model and reducing the false alarm rate of faults.

Citation Information

Patent Citations

  • Lithium ion battery pack voltage abnormity fault diagnosis method based on double models

    CN113687251A

  • Voltage prediction method and device for all-vanadium redox flow battery and medium

    CN115308608A

  • Lithium ion battery remaining service life prediction method based on algorithm fusion

    CN116794516A

  • CNN-GRU-based fuel cell voltage prediction method and device

    CN118376923A

Cited By

  • Flow battery voltage inflection point SOC verification method based on GBDT machine learning prediction

    CN119808681A

  • A method for verifying the voltage inflection point (SOC) of flow batteries based on GBDT machine learning prediction.

    CN119808681B

  • Fusion prediction method for proton exchange membrane fuel cell under dynamic working condition

    CN119885531A