A composite denoising LSTM prediction method for supercapacitor backup power supply performance degradation law and remaining life

Abstract

The present invention proposes a composite denoising LSTM prediction method for the performance degradation law and remaining service life of supercapacitor backup power supplies, which includes the following five steps: (1) data acquisition, (2) SG smoothing filtering, (3) MPA-VMD decomposition and denoising, (4) LSTM prediction, and (5) evaluation. The beneficial effects of the present invention are as follows: capacity is selected as a performance indicator, the SG smoothing method (Savitzky‑Golay Smoothing method, SG) is used to eliminate the noise generated by the capacity drop and recovery during the supercapacitor charging and discharging process, and the Marine Predators Algorithm (Marine Predators Algorithm) is used to optimize the variational mode decomposition (VMD) parameters, denoise the data, and reconstruct the denoised capacity sequence. Finally, the LSTM is used to predict the performance degradation law (PDL) and remaining useful life (RUL) of the supercapacitor. This effectively solves the impact of noise generated by factors such as capacity regeneration, charge and discharge rate differences, internal temperature changes of the supercapacitor, chemical reactions, and external electromagnetic interference during the service life of the supercapacitor backup power supply on the prediction accuracy, significantly reduces the RMSE, improves the R², and has high prediction accuracy. It can accurately predict the PDL and RUL of the supercapacitor backup power supply performance, thereby improving the safety and reliability of wind turbines operating under severe wind conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of supercapacitors, and in particular to a composite denoising LSTM prediction method for supercapacitor backup power supply performance degradation law and remaining life. Background Art

[0002] Wind power generation has become an important means of addressing global energy challenges, playing a crucial role in environmental protection and resource conservation. The pitch system, a crucial component of a wind turbine, absorbs wind energy and delivers rated power by adjusting the blade angle, maintaining safe turbine operation. A backup power supply, a key component of the pitch system, provides power. Supercapacitors, as exceptional energy storage devices, offer advantages such as high power density, rapid charge and discharge speeds, long cycle life, environmentally friendly disposal, and the ability to withstand high instantaneous charge and discharge currents. They are gradually replacing traditional batteries such as lead-acid batteries as the backup power source for wind turbine pitch systems. With the rapid maturity and widespread adoption of supercapacitor energy storage technology, the operational safety of supercapacitors as standalone or auxiliary energy storage systems is gaining increasing attention. The backup power supply typically operates intermittently. Normally, the pitch system relies on the grid for power, leaving the backup power source inactive. However, in extreme situations such as strong winds, severe weather, or a grid outage, the backup power source is activated to provide the power required for emergency feathering of the pitch system, ensuring safe and stable operation of the wind turbine. Therefore, how to accurately predict the remaining service life of supercapacitors and replace them in time before reaching the life threshold plays an important role in ensuring the normal operation of wind turbines and improving the adaptability and operational safety of wind turbines.

[0003] Existing methods for predicting the remaining service life of supercapacitors are divided into two categories: model-based methods and data-driven methods. Model-based methods aim to use parameter data such as voltage, current, and temperature, and adopt corresponding estimation methods to estimate the resistance, capacitance, and remaining margin of the supercapacitor equivalent circuit. These parameters are used as indicators, and then a physical mathematical model describing the aging behavior of the supercapacitor is established, and the remaining service life is predicted using this model. However, due to the high complexity, multivariate, and strong coupling characteristics of the electrochemical system inside the supercapacitor, it is difficult to establish an accurate mathematical model of aging behavior that adapts to the actual operating environment, resulting in low prediction accuracy. Data-driven methods do not need to consider specific physical mechanisms, but directly learn the common laws behind a large amount of data and update the data-driven model in real time based on the training data set. They have the advantages of simple structure, low complexity, and small estimation error.

[0004] However, due to the influence of factors such as electromagnetic interference, difference in charge and discharge rate, self-heating and chemical reaction in the supercapacitor during the whole life cycle charge and discharge process, there are abnormal capacity fluctuations and fluctuations in the aging curve of the capacity, and some irregular noise signals are mixed in. These interferences will affect the accuracy of supercapacitor life prediction; and in the early stage of supercapacitor life, the instability of its electrochemical reaction and charge transfer and the charging activation phenomenon will cause large fluctuations in capacity; in the middle of the supercapacitor life, the temperature change of the supercapacitor itself may cause the diffusion of the electrolyte and the change of the electrode reaction rate, resulting in large fluctuations in capacity, mixed with small fluctuations and high-frequency noise. When the life of a supercapacitor approaches a threshold, the performance of the supercapacitor degrades severely, and its capacity decreases rapidly with large fluctuations. Therefore, in order to improve the prediction accuracy of supercapacitor performance degradation and remaining service life, a supercapacitor performance degradation law (PLD) and remaining service life (RUL) LSTM prediction method based on composite denoising is invented. According to the different characteristics of the data of each life period in the entire life cycle of the supercapacitor, the composite denoising method is used to effectively reduce the noise, and the LSTM is used to accurately predict the performance degradation law and remaining service life of the supercapacitor. It is very necessary to replace the supercapacitor in time before it reaches the failure life threshold, so as to effectively improve the safety and reliability of wind turbines operating under severe wind conditions. Summary of the Invention

[0005] Aiming at the shortcomings of current supercapacitor performance degradation law and remaining service life prediction methods, the present invention discloses a composite denoising LSTM prediction method for supercapacitor backup power supply performance degradation law and remaining service life.

[0006] The present invention adopts the following technical solutions:

[0007] Step 1: Simulate the intermittent working mode of the supercapacitor backup power supply and perform an aging cycle test on the supercapacitor to obtain capacity data;

[0008] The aging cycle test steps of the supercapacitor are as follows:

[0009] 1) Use the battery tester model EBC-A10H to conduct charge and discharge cycle life tests, and conduct accelerated aging experiments on supercapacitors at a constant temperature of 65°C;

[0010] 2) Charge at a constant voltage until the charging current drops to the charging cut-off current, indicating that charging is complete;

[0011] 3) Discharge at a constant current until the supercapacitor voltage drops to the discharge cut-off voltage and the discharge process ends;

[0012] 4) After the charging and discharging process, set the rest time to allow the terminal voltage to stabilize.

[0013] Step 2: Use the SG smoothing filter algorithm to reduce the noise of the small fluctuations and high-frequency noise data in the original capacity data of the supercapacitor in the middle of its life, as follows:

[0014] Select original data x i There are M sample points on each side, and x i As the origin, construct an array with a window size of 2M+1, so that p polynomial to q ( n ) polynomial to fit this array:

[0015]

[0016] Where a k For the k The fitting coefficient of order;

[0017] After least square fitting, the residual C is

[0018]

[0019] Where x(n) is the data set to be fitted; when the residual C When it is the smallest, it means the fitting effect is the best;

[0020] Step 3: For the data segments with large fluctuations and complex noise in the original capacity data at the early, middle and near life threshold of the supercapacitor, the MPA-VMD ocean predator algorithm is used to optimize the variational mode decomposition method for denoising, as follows:

[0021] (1) Establishing the fitness function of MPA optimization algorithm f , as follows:

[0022] Weighted mean square error MSE and correlation coefficient r Constructing the fitness function for MPA optimization f for

[0023]

[0024] Where, w 1 and w 2 is the weight coefficient, the value range is [0,1], and w 1+ w 2=1; 、 ,in y i represents the original capacity sequence, x k (i ) represents the Kth modal component, n Indicates the number of samples; (2) Use MPA algorithm to optimize VMD parameters; First initialize MPA parameters, calculate and compare the current fitness F and the previous fitness value in the iteration F 0. If F < F 0, then F 0= F And calculate and update the predator position, in order to avoid the problem of local optimal solution, set the current number of iterations i Less than the maximum number of iterations i max ,until i > i max , obtain the global optimal fitness value and the top predator position, and thus obtain the optimal solution for the parameters K * 、 α * 、 β * 、 w 1 * 、 w 2 * ;

[0025] (3) According to the obtained K * 、 α * 、 β * , perform VMD variational mode decomposition denoising to obtain K * modal components;

[0026] Step 4: The maximum value of the modal component is greater than the threshold β The modal components of are superimposed to obtain the reconstruction capacity sequence:

[0027]

[0028] Where, u k ( t ) is the k modal components, max( u k ) indicates the k The maximum value of the modal component, (.) is the indicator function. When the conditions in the brackets are met, the indicator function value is 1, otherwise it is 0;

[0029] Step 5: Take the first 70% of the data in the reconstructed capacity sequence as the training set and the last 30% of the data as the prediction set. Use the LSTM neural network for training and prediction to obtain the predicted value of the remaining capacity and the remaining service life (RUL) of the supercapacitor. The change pattern of the remaining capacity predicted value reflects the performance degradation pattern of the supercapacitor. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the LSTM prediction method for the performance degradation law and remaining life of a supercapacitor backup power supply with composite denoising.

[0031] Figure 2 This is a graph showing the capacity data of the supercapacitor of the present invention.

[0032] Figure 3 This is the fourth section SG smoothing filter diagram of the present invention

[0033] Figure 4 This is the SG smoothing filter diagram of the fifth section of the present invention.

[0034] Figure 5 is the modal component diagram after VMD decomposition of the present invention

[0035] Figure 6 This is a comparison chart of the sequences before and after noise reduction of the present invention

[0036] Figure 7 This is the structural diagram of the LSTM of the present invention

[0037] Figure 8 This is the SG-VMD-LSTM prediction result diagram of the present invention

[0038] Figure 9 This is a comparison chart of predictions of different combination methods of the present invention Specific implementation plan

[0039] The solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0040] See attached figure, Figure 1 The present invention provides a flow chart of a composite denoising LSTM prediction method for supercapacitor backup power supply performance degradation law and remaining life. The method comprises the following five steps: (1) data acquisition, (2) SG smoothing filtering, (3) MPA-VMD decomposition and noise reduction, (4) LSTM prediction, and (5) evaluation.

[0041] Step 1: Data acquisition. Select a new supercapacitor with excellent performance as the calibration capacitor. The model is BCAP0350 E270 T11 350F produced by Maxwell, with a rated voltage of 2.7V.

[0042] A supercapacitor charge and discharge test platform was constructed, which included a battery tester and a data acquisition host computer. The battery tester was EBC-A10H, and the data acquisition host computer was a PC. Before the test, an electric heating constant temperature box was set to keep the experimental temperature constant at 65°C. Each supercapacitor cell was first discharged to 0.1V, and then the test was carried out. The test process included four stages: "charge-rest-discharge-rest": the first stage was the charging stage, which was first charged at a constant voltage of 2.7V, with the charging cut-off current set to 0.05A, until the charging current dropped to the charging cut-off current, indicating that the charging was completed; the second stage was the rest stage, in which the capacitor was left to rest for 5 minutes until the terminal voltage reached a stable level; the third stage was the discharge stage, which was discharged at a constant current of 3A, with the discharge cut-off voltage set to 0.1V, until the supercapacitor voltage dropped to the discharge cut-off voltage, and the discharge process was stopped; the fourth stage was the rest stage, which was left to rest for 5 minutes until the terminal voltage stabilized. According to the above four-stage test conditions of "charge-rest-discharge-rest", the supercapacitor is subjected to accelerated aging experiment on the test platform to obtain attribute data. The data includes total sampling time (s), charging current (A), discharge current (A), terminal voltage (V), capacity (Ah), number of cycles, etc. The sampling period is 2s.

[0043] See attached figure, Figure 2 This is the supercapacitor capacity data curve of the present invention. The ratio of the actual capacity value of the supercapacitor to the rated capacity value is used to measure the supercapacitor capacity, and is used as the supercapacitor capacity data for subsequent denoising and prediction. Its expression is:

[0044] ;

[0045] Where, C ( t ) is the supercapacitor in the first t The capacity value of the first charge and discharge cycle, C (0) is the rated capacity of the supercapacitor.

[0046] Step 2: SG smoothing filter; see the attached figure, Figure 3 This is the SG smoothing filter diagram of the fourth section of the present invention, Figure 4 This is the SG smoothing filter diagram for Section V of the present invention. The SG smoothing filter algorithm is used to reduce noise in the raw capacity data of supercapacitors during the mid-life period, including small fluctuations and high-frequency noise. A window of 2M+1 is selected as 50, and the smoothing order is set to 4. The SG smoothing method removes the fluctuations in the raw capacity data, resulting in a capacity series with a relatively stable degradation trend.

[0047] Step 3: MPA-VMD decomposition and noise reduction, for the data segments with large fluctuation and complex noise in the original capacity data at the early, middle and near life threshold of supercapacitor life, see the attached Figure 2 The Ⅰ, Ⅱ, and Ⅲ segments in the figure are used as VMD decomposition regions, and MPA is used to optimize the VMD decomposition layer number K and penalty factor. α , and determining the reconstruction threshold of the reconstructed signal β The optimal solution is to set the correlation coefficient r Combined with MSE as the fitness function, the global optimal fitness value and the top predator position are obtained, thus obtaining the parameters K 、 α 、 β 、 w 1. w 2 are 8, 620, 11, 0.3, and 0.7 respectively. The capacity sequence is decomposed into 8 intrinsic mode components (IMFs) through VMD. After VMD decomposition, IMF1-IMF8 components are obtained. The decomposed components are shown in the attached Figure 5 , Figure 5 The modal component diagram after VMD decomposition of the present invention is then analyzed using the optimal threshold β Select the effective components for reconstruction to obtain the denoised signal reflecting the degradation trend S ( t ), see Appendix Figure 6 , Figure 6 This is a comparison chart of the capacity sequence before and after noise reduction of the present invention.

[0048] Step 4: LSTM prediction; see attached figure, Figure 7 The LSTM structure diagram is used for training and prediction using the LSTM neural network model to obtain the predicted value of the remaining capacity and the remaining service life (RUL) of the supercapacitor. The change pattern of the remaining capacity predicted value reflects the performance degradation pattern of the supercapacitor.

[0049] See attached figure, Figure 8 This is the SG-VMD-LSTM prediction result diagram of the present invention. The capacity prediction curve selects the reaction degradation trend signal after noise reduction. S ( t) constitutes a prediction training set. The first 70% of the data in the capacity time series is taken as the training set, and the last 30% of the data is taken as the prediction set. The LSTM neural network is used for training and prediction to obtain the capacity prediction curve. The parameters of the LSTM neural network prediction model are set as follows: the input layer is 20; the hidden layer is 10; the Relu activation layer is 1; the fully connected layer is 1 and the learning rate is 0.005; the training cycle is 300 rounds, and each round of iteration is 12 times. After the training iteration, the capacity of the supercapacitor corresponding to the target position is predicted. When the prediction begins, the number of charge and discharge cycles corresponding to the starting point is T0=1604. When the capacity prediction value reaches the failure threshold (80%), the corresponding number of supercapacitor charge and discharge cycles T0=2201 is recorded. The supercapacitor RUL is calculated to be 597 charge and discharge cycles.

[0050] Step 5: Evaluation: The prediction of performance degradation law (PDL) and remaining useful life (RUL) is evaluated. The advantages and effectiveness of this method are verified through different combinations of algorithms and comparative analysis with other prediction methods.

[0051] The following section verifies the effectiveness of the composite denoising LSTM prediction method for supercapacitor backup power supply performance degradation patterns and remaining life proposed in this patent. The capacity data used was obtained through accelerated aging experiments at a constant temperature of 65°C. When the supercapacitor capacity falls below 80%, it is considered to have reached the failure threshold. The prediction target is the number of charge and discharge cycles when the capacity curve drops to 80%. The experimental scenario uses the first 70% of the capacity time series data as the training set for the LSTM neural network training, and the last 30% of the capacity time series data as the prediction set for LSTM prediction of PDL and RUL.

[0052] See attached figure, Figure 9The prediction comparison diagram of different combination methods of the present invention shows that the SG-LSTM curve is at the top and deviates the most from the experimental curve. The SG-LSTM method uses SG smoothing filtering to preprocess the entire raw data, and SG smoothing filtering is suitable for removing local small fluctuations and high-frequency noise. It is too smooth for the large fluctuations in capacity caused by the charging activation phenomenon in the early charge and discharge cycle of the supercapacitor, which seriously affects the LSTM network training accuracy and results in the worst performance degradation prediction accuracy. The VMD-LSTM curve is at the top, deviates from the experimental curve secondly and has a certain degree of fluctuation, especially when approaching the failure threshold. The fluctuation is large because the VMD-LSTM method uses VMD modal decomposition and denoising to preprocess the entire raw data, and VMD modal decomposition and denoising is suitable for the nonlinear and non-stationary noise of capacity data in the supercapacitor cyclic charge and discharge process. It over-decomposes and denoises the stationary region of the supercapacitor mid-term charge and discharge cycle, affecting the LSTM network training accuracy and resulting in poor performance degradation prediction accuracy. The LSTM curve is below the experimental curve, with a small deviation from the experimental curve, but exhibits sawtooth ripples. This is mainly because the LSTM method uses raw data without any denoising during network training and prediction. The noise affects the LSTM network training accuracy, resulting in low performance degradation prediction accuracy. The curve of the composite denoising supercapacitor backup power supply performance degradation law and remaining life LSTM prediction method proposed in this invention is below the experimental curve, smooth, and has minimal deviation from the experimental curve. This is mainly because the method proposed in this article targets the noise generated by the characteristics of different stages in the supercapacitor aging process, and adopts SG smoothing filtering and VMD modal decomposition denoising methods to segmentally denoise the raw data of the supercapacitor in the early, middle, and near life threshold stages, effectively removing noise interference and significantly improving the LSTM network training accuracy and performance degradation prediction accuracy. Table 2 shows that the RMSEs for the SG-VMD-LSTM, VMD-LSTM, SG-LSTM, and LSTM methods over the entire prediction phase are 1.655, 2.898, 3.850, and 3.547, respectively. SG-VMD-LSTM has the smallest RMSE, while the RMSEs for the other three methods are 1.75, 2.33, and 2.14 times that of our method, respectively. The R² values ​​for these methods are 0.955, 0.863, 0.759, and 0.796, respectively. The R² for SG-VMD-LSTM is greater than 0.95, closest to 1. The R² values ​​for the other three methods are 9.6%, 20.5%, and 16.6% lower than those for our method, respectively. This shows that our method has the highest accuracy for predicting supercapacitor performance degradation trends.

[0053]

[0054] Experimental measurements show that the supercapacitor's true RUL is 597 charge-discharge cycles. Table 4 shows that the RULs for the SG-VMD-LSTM, VMD-LSTM, SG-LSTM, and LSTM methods are 603, 620, 622, and 623, respectively, with RUL errors of 0.84%, 3.85%, 4.19%, and 4.35%, respectively. The proposed method has the smallest RUL error, while the RUL errors of the other methods are approximately five times greater. The backup power supply typically operates intermittently. Normally, the pitch system is powered by the grid, and the backup power supply is idle. It is only activated in extreme situations, such as strong winds, severe weather, or grid outages, and its usage frequency is low. Furthermore, the closer a supercapacitor is to its failure threshold, the longer its charge-discharge cycle takes. However, since the number of supercapacitor charge-discharge cycles is used as the RUL, even a slight prediction error can lead to significant errors in the remaining service life. Therefore, in view of the different types of noise caused by the different characteristics of the supercapacitor backup power supply at different stages of its service life, this patent proposes a composite denoising supercapacitor backup power supply performance degradation law and remaining life LSTM prediction method. The present invention can more accurately predict the backup power supply RUL, and can timely and accurately grasp the performance degradation law and remaining service life of the supercapacitor backup power supply. The backup power supply of the wind turbine can be replaced in time before it reaches the life threshold, thereby improving the safety and reliability of the wind turbine operation under severe wind conditions.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A supercapacitor performance degradation law and remaining service life LSTM prediction method based on composite denoising. First, the capacity data of the supercapacitor during the aging cycle is obtained through a supercapacitor charge and discharge test platform. Different denoising methods are used to perform composite denoising on the data of different life periods of the supercapacitor, and the capacity data sequence is reconstructed. Then, LSTM is used to predict the supercapacitor performance degradation law and remaining service life RUL. The RUL prediction value is ,in The number of cycles at which the supercapacitor's life is predicted to end, The cycle number where the supercapacitor predicts the starting point is RUL error , root mean square error 、 Evaluate the prediction results, n To predict the number of cycles, is the average value of the capacity test value, x i is the capacity experimental value, is the capacity prediction value; it is characterized by: A supercapacitor performance degradation law and remaining life LSTM prediction method based on composite denoising is carried out in the following steps: Step 1: Use the SG smoothing filter algorithm to reduce noise in the small fluctuations and high-frequency noise data in the original capacity data of the supercapacitor in the middle of its life, as follows: Select original data x i There are M sample points on each side, and x i As the origin, construct an array with a window size of 2M+1, so that p Order polynomial q ( n ) polynomial to fit this array: （1）； Where a k For the k The fitting coefficient of order; After least square fitting, the residual C is （2）； Where x(n) is the data set to be fitted; when the residual C When it is the smallest, it means the fitting effect is the best; Step 2: For the data segments with large fluctuations and complex noise in the original capacity data at the early, middle and near life threshold of the supercapacitor, the MPA-VMD ocean predator algorithm is used to optimize the variational mode decomposition method for denoising, as follows: (1) Establishing the fitness function of MPA optimization algorithm f , as follows: Weighted mean square error MSE and correlation coefficient r Constructing the fitness function for MPA optimization f for （3）； Where, w 1 and w 2 is the weight coefficient, the value range is [0,1], and w 1+ w 2=1; 、 ,in y i represents the original capacity sequence, x k ( i ) represents the Kth modal component, n Indicates the number of samples; (2) Using the MPA algorithm to optimize the VMD parameters, the global optimal fitness value and the top predator position will be obtained, thereby obtaining the VMD decomposition layer K and the penalty factor α , threshold β , weight coefficient w 1 and w The optimal solution of 2 K * 、 α * 、 β * 、 w 1 * 、 w 2 * ; (3) According to the obtained K * 、 α * 、 β * , perform VMD variational mode decomposition denoising to obtain K * modal components; Step 3: The maximum value of the modal component is greater than the threshold β The modal components of are superimposed to obtain the reconstruction capacity sequence: （4）； Where, u k ( t ) is the k modal components, max( u k ) indicates the k The maximum value of the modal components, is an indicator function. When the condition in the brackets is met, the indicator function value is 1, otherwise it is 0; Step 4: Take the first 70% of the data in the reconstructed capacity sequence as the training set and the last 30% of the data as the prediction set. Use the LSTM neural network for training and prediction to obtain the predicted value of the remaining capacity and the remaining service life (RUL) of the supercapacitor. The change pattern of the remaining capacity predicted value reflects the performance degradation pattern of the supercapacitor.

Images