All-vanadium redox flow battery pump fault diagnosis method based on optimized LSTM

Abstract

The invention relates to an all-vanadium redox flow battery pump fault diagnosis method based on optimized LSTM. The method does not depend on a high-precision sensor, characteristic parameters such as current, voltage, flow velocity and SOC are extracted from historical data of the battery, parameters of the long-short-term memory network are optimized, finally, a fault diagnosis result is obtained through training of the long-short-term memory network, and diagnosis and classification of faults of the positive electrode side pump, the negative electrode side pump and the double-side pump of the all-vanadium redox flow battery are completed. The artificial rabbit algorithm is adopted to optimize and solve the LSTM hyper-parameter problem, and the accuracy of fault diagnosis and classification of the all-vanadium redox flow battery pump can be effectively improved.

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 pump fault detection method based on optimized LSTM. Background Art

[0002] As the application market's demand for energy storage system reliability and capacity stability increases year by year, the deployment speed of energy storage systems for grid applications is accelerating. As we all know, electrochemical energy storage technology is a key link in the development of fixed power energy storage systems. Among the many electrochemical energy storage technologies, all-vanadium liquid flow batteries have extensive commercial application value due to their high safety, independent output power and capacity, strong overload capacity, strong deep discharge capability and long cycle life. Battery fault diagnosis plays an increasingly important role in modern industrial systems. With the rapid development of big data processing technology, data-driven methods have been increasingly widely used in the field of energy storage battery fault diagnosis.

[0003] In existing research, the LSTM method is mainly used for fault diagnosis of lithium batteries, while research on all-vanadium liquid flow batteries mostly focuses on battery performance analysis, electrode material optimization and flow path optimization, and few studies focus on fault diagnosis of all-vanadium liquid flow batteries. Summary of the invention

[0004] The purpose of the present invention is to provide a vanadium liquid flow battery pump fault detection method based on optimized LSTM, and the LSTM method is applied to the fault diagnosis of the vanadium liquid flow battery. The method does not require any flow sensor and can classify the positive pump fault, negative pump fault and double-side pump fault of the vanadium liquid flow battery pump. The LSTMLSTM algorithm is used to detect the operating status of the vanadium liquid flow battery water pump in actual operation, and the parameters in the LSTM are optimized by ARO, which can well solve the technical problem of vanadium liquid flow battery pump fault detection.

[0005] To achieve the above purpose, the present invention designs a vanadium liquid flow battery pump fault detection method based on optimized LSTM, comprising the following steps:

[0006] Step 1: Conduct simulated fault experiments on the positive electrode pump, negative electrode pump and double-side pump of the all-vanadium liquid flow battery. The specific method of the simulated fault experiment is as follows: the pump is stopped several times during the charge and discharge process. The pump stop process is regarded as a pump failure, and the total charge and discharge time is greater than the total pump stop time. After that, the battery is fully charged or fully discharged. In the simulated fault experiment, the experimental data of the all-vanadium liquid flow battery under the conditions of positive electrode pump failure, negative electrode pump failure or double-side pump failure are recorded;

[0007] Step 2: Extract the voltage and current data of the battery from the experimental data of step 1 as characteristic parameters 1 and 2. When a pump failure occurs, introduce the rate of change of the terminal voltage at different currents shown in formula (1) as characteristic parameter 3:

[0008]

[0009] Among them, ΔU represents the voltage difference between two moments, and I represents the current at a certain moment;

[0010] Step 3: Introduce the intrinsic relationship between the state of charge SOC and current, as shown in formula (2):

[0011]

[0012] Among them, SOC0 is the initial SOC of the battery, I is the charge and discharge current, C n is the rated capacity of the battery;

[0013] Step 4: Use a small current to perform a charge and discharge operation on the battery, obtain the battery's charge and discharge OCV-SOC curve based on the SOC calculation method in step 3, obtain the battery's OCV-SOC curve by calculating the average of the two curves, and solve the corresponding SOC value as characteristic parameter 4 based on the voltage data obtained in step 2;

[0014] Step 5: Use principal component analysis PCA to reduce the dimensionality of the four-dimensional feature parameters composed of the above feature parameters 1 to 4, and calculate the cumulative variance contribution rate μ of the eigenvalues ​​of the real symmetric covariance matrix e To determine whether the obtained principal component effectively summarizes most of the information of the four-dimensional feature parameters, the calculation is as follows:

[0015]

[0016] Among them, p is the dimension of the input feature parameter, is the eigenvalue e j The variance contribution rate of

[0017] Step 6: When the variance contribution rate μ obtained in step 5 e When it reaches 80% or higher, it means that the selected eigenvalues ​​can effectively summarize the main information of the four characteristic parameter sample data sets; if the first r sample points of the sample data are selected as the principal components, the remaining sample vectors are calculated and transformed into the residual matrix M r :

[0018]

[0019] The matrix M is used to normalize the matrix of the initial high-dimensional data to be processed, and the standardized matrix obtained by the normalization calculation is reduced in dimension to obtain the reduced n-sample length r-dimensional matrix S n×r ;in It is the n-sample length r-dimensional matrix before dimensionality reduction.

[0020]

[0021] Step 7: Use 70% of the principal component data as the training set of the pump fault diagnosis model, initialize the rabbit population, and calculate the energy factor. Obtain the optimal position information by updating the circuitous foraging position information and random hidden position information under different energy factor conditions. Finally, determine whether to output the optimal LSTM network parameters by judging whether the termination condition is met. The optimal LSTM network parameters include hidden layer ls, maximum training times ep, initial learning rate lr, learning rate reduction factor lrdf, and learning rate reduction period lrdp;

[0022] Step 8: According to the optimal LSTM network parameters obtained in step 7, an all-vanadium liquid flow battery pump fault classifier based on the optimized LSTM is established, and the remaining 30% of the principal component data is used as a test set to output the pump fault classification result.

[0023] Furthermore, in step 1, the charging process of the simulated fault experiment is specifically as follows: the pump is stopped twice during the charging process, first charging for 4 minutes and stopping the pump for 30 seconds; then charging for another 4 minutes and stopping the pump for another 30 seconds; and then fully charging the battery.

[0024] Furthermore, in step 7, the optimal position information is obtained by updating the circuitous foraging position information and the random hidden position information under different energy factors, which specifically includes the following steps:

[0025] Step 701: During each iteration of the algorithm, the rabbit determines its current behavior based on the value of the energy factor A(t). The expression of the energy factor A(t) is as follows:

[0026]

[0027] Where r0 represents a random number in the range of (0,1), t represents the number of iterations, and T represents the maximum number of iterations;

[0028] Step 702: A(t) is calculated. If the energy factor A(t) is greater than 1, the rabbit performs a circuitous foraging behavior. If the energy factor A(t) is less than or equal to 1, the rabbit performs a random hiding behavior. As the number of iterations increases, the value of A(t) gradually decreases.

[0029] Step 703: By introducing disturbance elements to simulate the behavior of the rabbit when foraging, the algorithm finally finds the approximate optimal solution to the optimization problem.

[0030] Furthermore, in step 6, the sample data is the number of all-vanadium liquid flow batteries, and the data categories include actual classification and predicted classification. If the actual classification and the predicted classification symbols coincide, the classification is correct, and if they do not coincide, the classification is incorrect. The calculation method of the predicted classification accuracy is (number of correctly classified samples / total number of samples)*100%.

[0031] Beneficial effects of the present invention:

[0032] The present invention applies the LSTM algorithm optimized based on the artificial rabbits optimization algorithm (ARO) to the fault diagnosis of the all-vanadium liquid flow battery. Because ARO is a method of optimizing the algorithm, it is widely used in the LSTM algorithm. The method is applicable to the power grid energy storage system. The method does not require any flow sensor and can classify the positive electrode pump fault, negative electrode pump fault and double-side pump fault of the all-vanadium liquid flow battery pump. The characteristic parameters are extracted from the battery state curve, the parameters of the LSTM are optimized, and finally the fault prediction results are obtained through LSTM network training to realize the fault classification of the positive electrode pump fault, negative electrode pump fault and double-side pump fault of the all-vanadium liquid flow battery pump. The LSTM algorithm is used to detect the operating state of the all-vanadium liquid flow battery water pump in actual operation, and the parameters in the LSTM are optimized by the ARO algorithm, which can well solve the technical problem of all-vanadium liquid flow battery pump fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a method for detecting a fault of an all-vanadium liquid flow battery pump based on optimized LSTM according to an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of parameter optimization results of the present invention;

[0035] Figure 3 This is a schematic diagram of the prediction classification of the present invention. DETAILED DESCRIPTION

[0036] In order to make the technical problems solved by the present invention, the technical solutions adopted and the technical effects achieved clearer, the technical solutions of the present invention are further described below in conjunction with the accompanying drawings and through specific implementation methods. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, not all.

[0037] like Figure 1 As shown, the embodiment of the present invention proposes a method for detecting a fault of an all-vanadium liquid flow battery pump based on an optimized LSTM, comprising:

[0038] Step 1: Conduct simulated fault experiments on the positive electrode pump, negative electrode pump and double-side pump of the all-vanadium liquid flow battery:

[0039] (i) First use the current I t Fully discharge the battery;

[0040] (ii) Reuse current I t Charge the battery;

[0041] (iii) turning off the positive electrode pump to simulate a positive electrode pump failure; and continuing to charge the battery until the battery is fully charged;

[0042] (iv) turning off the cathode pump to simulate a cathode pump failure; and continuing to charge the battery until the battery is fully charged;

[0043] (v) turning off the pumps on both sides at the same time to simulate a failure of the pumps on both sides; and continuing to charge the battery until the battery is fully charged;

[0044] (vi) Record the experimental data of the all-vanadium liquid flow battery under the conditions of positive electrode pump failure, negative electrode pump failure or double-side pump failure;

[0045] The experimental parameter settings of the all-vanadium liquid flow battery in this example are shown in Table 1. The pump is stopped twice during the charging process. First, the pump is charged for 4 minutes and then stopped for 30 seconds; then the battery is fully charged for 4 minutes and then stopped for 30 seconds. The discharge process is exactly the opposite of the charging process, so it will not be repeated. In the all-vanadium liquid flow battery pump failure simulation experiment, the battery voltage increased significantly during the pump stop process during charging; and the battery voltage decreased significantly during the pump stop process during discharge.

[0046] Table 1 Experimental parameter settings of all-vanadium liquid flow battery

[0047]

[0048]

[0049] Extract the change curve of current and voltage signals from the experimental data of step 1. Since the voltage signal is affected by the electrochemical properties, temperature, flow rate and material properties of the all-vanadium liquid flow battery. Therefore, the voltage signal is one of the key signals in the health detection indicators of the all-vanadium liquid flow battery. In the event of a fault, the change in the voltage signal is the most direct and obvious. Because current changes will also cause voltage changes, there is internal resistance in the battery, and the current increases, the voltage drop on the resistance will increase, causing the overall voltage to increase. Therefore, in order to eliminate the influence of current changes on voltage, the resistance change rate ΔR is introduced to intuitively reflect the state of the battery, and is expressed by formula (1):

[0050]

[0051] Among them Ut is the voltage value collected at a certain moment; U t-1 is the voltage value collected within one second before a certain moment; I t It is the current value collected at a certain moment.

[0052] Step 2: Extract the battery voltage and current data from the experimental data in step 1 as characteristic parameters 1 and 2. When a pump failure occurs, the voltage changes most significantly. To prevent the terminal voltage from changing suddenly due to the constant current change under complex working conditions, the rate of change of the terminal voltage at different currents shown in formula (1) is introduced as characteristic parameter 3.

[0053]

[0054] Among them, ΔU represents the voltage difference between two moments, and I represents the current at a certain moment.

[0055] Step 3: Introduce the intrinsic relationship between the state of charge (SOC) and the battery terminal voltage, as shown in the following equation.

[0056]

[0057] Among them, SOC0 is the initial SOC of the battery, I is the charge and discharge current, C n is the rated capacity of the battery.

[0058] Step 4: Use a small current to perform a charge and discharge operation on the battery, obtain the battery's charge and discharge OCV-SOC curve based on the SOC calculation method in step 3, and obtain the battery's OCV-SOC curve by calculating the average of the two curves. According to the voltage data obtained in step 2, solve the corresponding SOC value as characteristic parameter 4.

[0059] Step 5: To reduce the model load and computing time, principal component analysis (PCA) is used to reduce the dimensionality of the four-dimensional feature parameters composed of the above feature parameters 1 to 4. The cumulative variance contribution rate μ of the eigenvalues ​​of the real symmetric covariance matrix is ​​calculated. e To determine whether the obtained principal components effectively summarize most of the information of the four-dimensional feature parameters, the calculation is shown in the following formula.

[0060]

[0061] Among them, p is the dimension of the input feature parameter, is the eigenvalue e j The variance contribution rate of .

[0062] Step 6: When the variance contribution rate μ obtained in step 5 eWhen it reaches 80% or higher, it means that the selected eigenvalues ​​can effectively summarize the main information of the four characteristic parameter sample data sets; if the first r sample points of the sample data are selected as the principal components, the remaining sample vectors are calculated and transformed into the residual matrix M r :

[0063]

[0064] The matrix M is used to normalize the matrix of the initial high-dimensional data to be processed, and the standardized matrix obtained by the normalization calculation is reduced in dimension to obtain the reduced n-sample length r-dimensional matrix S n×r .in It is the n-sample length r-dimensional matrix before dimensionality reduction.

[0065]

[0066] Step 7: Based on step 6, the principal component data representing the four characteristic parameters are obtained, and 70% of the principal component data are used as the training set of the pump fault diagnosis model. Initialize the rabbit population and calculate the energy factor. Obtain the optimal position information by updating the circuitous foraging position information and random hidden position information under different energy factor conditions. Finally, by judging whether the termination condition is met, decide whether to output the optimal network parameters hidden layer ls, maximum number of training times ep, initial learning rate lr, learning rate reduction factor lrdf, learning rate reduction period lrdp (such as Figure 2 shown).

[0067] Step 8: Establish an optimized LSTM-based all-vanadium liquid flow battery pump fault classifier according to the optimal LSTM network parameters obtained in step 7. Use the remaining 30% of the principal component data as a test set to output the pump fault classification results.

[0068] Wherein, the step 7 includes introducing an energy factor to describe and balance the global search and local optimization to prevent falling into a local extreme value.

[0069] Among them, step 7 obtains the optimal position information by updating the circuitous foraging position information and the random hidden position information under different energy factor conditions, which specifically includes the following steps:

[0070] Step 701: During each iteration of the algorithm, the rabbit will determine its current behavior based on the value of the energy factor A(t). The expression of the energy factor A(t) is as follows:

[0071]

[0072] Where r0 represents a random number in the range (0,1), t represents the number of iterations, and T represents the maximum number of iterations.

[0073] Step 702: Calculate A(t). If the energy factor A(t) is greater than 1, the rabbit will perform a circuitous foraging behavior. If the energy factor A(t) is less than or equal to 1, the rabbit will perform a random hiding behavior. As the number of iterations increases, the value of A(t) gradually decreases. This design forces each individual in the population to alternate between circuitous foraging behavior and random hiding behavior to balance the needs of global search and local optimization.

[0074] Step 703: By introducing disturbance elements to simulate the behavior of rabbits when foraging, this disturbance mechanism helps the artificial rabbit algorithm avoid falling into local extremes and enhances its search ability in the global scope. Finally, the algorithm finds the approximate optimal solution to the optimization problem.

[0075] Figure 3 The figure is a schematic diagram of the classification accuracy rate in the case of a positive electrode pump failure. The horizontal axis in the figure is the sample data x, and the sample data 0-200 represent the sample number, that is, the number of all-vanadium liquid flow battery failure samples. The vertical axis is the fault category; it can be seen from the figure that when the peristaltic pump is in normal operation, the overlap rate is 100%; when the peristaltic pump is in a faulty state, there is a partial non-overlap.

[0076] The calculation method for the prediction classification accuracy is: (number of correctly classified samples / total number of samples)*100%. Input the experimental parameters and according to the LSTM algorithm, the classification accuracy in the case of positive pump failure is 96.88%. The classification accuracy diagram in the case of double-side pump failure and negative pump failure is shown in Figure 2. Figure 3 Similarly, according to the LSTM algorithm, the classification accuracy rates in the case of double-side pump failure and negative electrode pump failure are 97.22% and 93.71% respectively.

[0077] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A fault diagnosis method for all-vanadium liquid flow battery pump based on optimized LSTM, characterized in that: The following steps are involved: Step 1: Conduct simulated fault experiments on the positive electrode pump, negative electrode pump and double-side pump of the all-vanadium liquid flow battery. The specific method of the simulated fault experiment is as follows: the pump is stopped several times during the charge and discharge process. The pump stop process is regarded as a pump failure, and the total charge and discharge time is greater than the total pump stop time. After that, the battery is fully charged or fully discharged. In the simulated fault experiment, the experimental data of the all-vanadium liquid flow battery under the conditions of positive electrode pump failure, negative electrode pump failure or double-side pump failure are recorded; Step 2: Extract the voltage and current data of the battery from the experimental data of step 1 as characteristic parameters 1 and 2. When a pump failure occurs, introduce the rate of change of the terminal voltage at different currents shown in formula (1) as characteristic parameter 3: Among them, ΔU represents the voltage difference between two moments, and I represents the current at a certain moment; Step 3: Introduce the intrinsic relationship between the state of charge SOC and current, as shown in formula (2): Among them, SOC0 is the initial SOC of the battery, I is the charge and discharge current, C n is the rated capacity of the battery; Step 4: Use a small current to perform a charge and discharge operation on the battery, obtain the battery's charge and discharge OCV-SOC curve based on the SOC calculation method in step 3, obtain the battery's OCV-SOC curve by calculating the average of the two curves, and solve the corresponding SOC value as characteristic parameter 4 based on the voltage data obtained in step 2; Step 5: Use principal component analysis PCA to reduce the dimensionality of the four-dimensional feature parameters composed of the above feature parameters 1 to 4, and calculate the cumulative variance contribution rate μ of the eigenvalues ​​of the real symmetric covariance matrix e To determine whether the obtained principal component effectively summarizes most of the information of the four-dimensional feature parameters, the calculation is as follows: Among them, p is the dimension of the input feature parameter, is the eigenvalue e j The variance contribution rate of Step 6: When the variance contribution rate μ obtained in step 5 e When it reaches 80% or higher, it means that the selected eigenvalues ​​can effectively summarize the main information of the four characteristic parameter sample data sets; if the first r sample points of the sample data are selected as the principal components, the remaining sample vectors are calculated and transformed into the residual matrix M r : The matrix M is used to normalize the matrix of the initial high-dimensional data to be processed, and the standardized matrix obtained by the normalization calculation is reduced in dimension to obtain the reduced n-sample length r-dimensional matrix S n×r ;in It is the n-sample length r-dimensional matrix before dimensionality reduction. Step 7: Use 70% of the principal component data as the training set of the pump fault diagnosis model, initialize the rabbit population, and calculate the energy factor. Obtain the optimal position information by updating the circuitous foraging position information and random hidden position information under different energy factor conditions. Finally, determine whether to output the optimal LSTM network parameters by judging whether the termination condition is met. The optimal LSTM network parameters include hidden layer ls, maximum training times ep, initial learning rate lr, learning rate reduction factor lrdf, and learning rate reduction period lrdp; Step 8: According to the optimal LSTM network parameters obtained in step 7, an all-vanadium liquid flow battery pump fault classifier based on the optimized LSTM is established, and the remaining 30% of the principal component data is used as a test set to output the pump fault classification result.

2. The all-vanadium liquid flow battery pump fault detection method based on optimized LSTM according to claim 1, characterized in that: In step 1, the charging process of the simulated fault experiment is specifically as follows: the pump is stopped twice during the charging process, first charging for 4 minutes and stopping the pump for 30 seconds; then charging for another 4 minutes and stopping the pump for another 30 seconds; and then fully charging the battery.

3. The all-vanadium liquid flow battery pump fault diagnosis method based on optimized LSTM according to claim 1, characterized in that: In step 7, the optimal position information is obtained by updating the circuitous foraging position information and the random hidden position information under different energy factors, which specifically includes the following steps: Step 701: During each iteration of the algorithm, the rabbit determines its current behavior based on the value of the energy factor A(t). The expression of the energy factor A(t) is as follows: Where r0 represents a random number in the range of (0,1), t represents the number of iterations, and T represents the maximum number of iterations; Step 702: A(t) is calculated. If the energy factor A(t) is greater than 1, the rabbit performs a circuitous foraging behavior. If the energy factor A(t) is less than or equal to 1, the rabbit performs a random hiding behavior. As the number of iterations increases, the value of A(t) gradually decreases. Step 703: By introducing disturbance elements to simulate the behavior of the rabbit when foraging, the algorithm finally finds the approximate optimal solution to the optimization problem.

4. The all-vanadium liquid flow battery pump fault diagnosis method based on optimized LSTM according to claim 1, characterized in that: The sample data in step 6 is the number of all-vanadium liquid flow batteries. The data categories include actual classification and predicted classification. If the actual classification and predicted classification symbols coincide, the classification is correct. If they do not coincide, the classification is incorrect. The calculation method for the predicted classification accuracy is (number of correctly classified samples / total number of samples)*100%.

Images