Photovoltaic array fault diagnosis method based on photovoltaic power station current signal output

Abstract

The invention discloses a photovoltaic array fault diagnosis method based on photovoltaic power station current signal output, and the method comprises the steps: collecting a current signal, carrying out the denoising of the current signal, and decomposing the denoised current signal into a plurality of intrinsic mode functions IMFs through employing an improved AVMD algorithm; adaptive learning rate adjustment is introduced into the improved AVMD algorithm, and bandwidth parameters and frequency parameters are optimized and updated in combination with the improved Adam algorithm; a Lasso regression model is constructed, L1 regularization is introduced, and IMF with high importance is screened out to serve as an information carrier for diagnosing photovoltaic system faults; wavelet transformation is carried out on the selected IMF, the IMF is decomposed into low-frequency components and high-frequency components on different frequency bands, coarse graining processing is carried out on the low-frequency components, then multi-scale discrete entropy (MDE) calculation is carried out, and a time scale factor and a normalized discrete entropy value curve are obtained; and setting a threshold value to judge the health condition of the photovoltaic system so as to realize fault diagnosis. The improved AVMD algorithm is utilized to decompose and denoise the current signal, and the method has good robustness and is easy to implement.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and relates to a photovoltaic array fault diagnosis method, and specifically to a photovoltaic array fault diagnosis method based on current signal output of a photovoltaic power station. Background Art

[0002] As the impact of industrial pollution on the environment gradually increases, the world has begun to pay attention to environmental protection, reduce the use of mineral fuels, and turn to renewable clean energy. Among them, solar energy, as one of the most productive renewable clean energy sources, has been widely disseminated and used. Regions with great solar energy potential have increased photovoltaic power generation and the number of solar power plants. With the increasing proportion of intermittent energy such as photovoltaics, this has undoubtedly increased the risk of drastic changes in energy production. Fault diagnosis can promptly detect potential problems in photovoltaic systems, so that measures can be taken to repair them and improve the operating efficiency and power generation of the system.

[0003] Some traditional photovoltaic degradation fault detection methods require the use of electroluminescence, infrared thermal imaging, ultraviolet fluorescence and IV tracers, which are costly, so many manufacturers will only use these methods when the power generation efficiency drops significantly. The DETECT (Online Diagnosis of Electronic System Health) research project was carried out in this context, with the aim of developing a low-cost FDD (photovoltaic system diagnosis) that does not require additional sensors or expensive equipment. The most commonly used FDD method is output signal analysis (OSA), which assumes that any fault will affect the output of the photovoltaic system, and the output current or voltage may be affected by numerical drops or dynamic changes in the signal. By fitting the disturbance, the pattern curve of each fault is extracted. However, OSA-based diagnosis is still sensitive to noise, so studying how to reduce the impact of noise on detection capabilities is a difficult problem that needs to be solved in output signal analysis. Summary of the invention

[0004] In order to determine the specific type of photovoltaic fault, the present invention provides a photovoltaic system fault diagnosis method based on an improved adaptive variational mode decomposition algorithm and multi-scale discrete entropy, which can perform online diagnosis based on the current signal output by the photovoltaic power station, so as to quickly diagnose the type of photovoltaic fault with low cost and easy operation.

[0005] To achieve the above object, the present invention provides the following technical solution: a photovoltaic array fault diagnosis method based on the current signal output of a photovoltaic power station, comprising the following steps:

[0006] Step 1. Connect the operation and management system of the photovoltaic power station to the inverter of the power grid to collect current signals, collect current signals of photovoltaic modules blocked by shadows and current signals in normal weather, use Kalman filtering to denoise the input current signals, and use the improved AVMD algorithm to decompose the denoised current signals into multiple intrinsic mode functions IMFs; the improved AVMD algorithm introduces adaptive learning rate adjustment, and combines the improved Adam algorithm to optimize and update the bandwidth parameters and frequency parameters;

[0007] Step 2. Apply the Lasso regression model to select the best regularization parameter using cross validation to balance the bias and variance of the model, fit the Lasso regression model using the training data, calculate the regression coefficient of the IMF, evaluate the importance of the IMF in fault diagnosis, and then determine the IMFs with higher importance as the information carrier for diagnosing photovoltaic system faults;

[0008] Step 3. Perform wavelet transform on the selected IMF to decompose the original time series of the output current of the photovoltaic power station into low-frequency components and high-frequency components in different frequency bands;

[0009] Step 4. Coarse-graining the low-frequency components, and then calculating the multi-scale discrete entropy MDE on the coarse-grained subsequences of different time scales, and normalizing the calculated discrete entropy values;

[0010] Step 5. Compare the normalized discrete entropy value curves of fault-free sunny days, fault-free cloudy days, and fault-free sunny days, observe the differences in discrete entropy values ​​at different time scales, find out the discrete entropy value with a large difference between the photovoltaic system with shading fault and the one without fault, set it as the threshold, and use it to judge whether the photovoltaic system is in a normal state, a shading fault state, and whether the shading fault occurs or ends, so as to realize fault diagnosis of the photovoltaic system.

[0011] Furthermore, in step 1, the specific method of decomposing the current signal output by the photovoltaic power station inverter into multiple intrinsic mode functions using the improved AVMD is:

[0012] For the current signal after Kalman filtering, the goal of improving AVMD is to decompose the current signal into K modal functions u k (t) and the corresponding center frequency ω k , the decomposition objective is to minimize the following bandwidth-constrained variational formula:

[0013]

[0014] in, represents the time derivative, which is used to represent the change of bandwidth, u k (t) is the kth mode function of the signal, ω kis the center frequency of each mode function, which is dynamically updated with adaptive learning rate and improved Adam optimization;

[0015] The Lagrangian multiplier λ(t) is introduced in the improved AVMD, so that the sum of the modal components can reconstruct the original signal, making the reconstructed signal as close to the original signal as possible to ensure the integrity of the signal. The Lagrangian function is defined as:

[0016]

[0017] Among them, α is the adaptive learning rate, which is used to dynamically control the optimization step size of the bandwidth, λ(t) is the Lagrange multiplier, which ensures the reconstruction accuracy of the signal, f(t) is the input original current signal of the photovoltaic power station, and K is the number of modal components, which represents the number of intrinsic mode functions obtained when decomposing the signal;

[0018] Introduce the improved Adam optimization algorithm to optimize u k (t) and ω k Dynamic update, first, Adam optimization algorithm initializes parameters, sets the initial learning rate and two exponential decay rate parameters β 1 and β 2 , initialize the first-order moment m 0 = 0 and the second moment v 0 =0, set a small value ∈ to avoid division by zero error;

[0019] The calculation steps of the improved Adam optimization algorithm are as follows:

[0020] a. Calculate the gradient

[0021]

[0022] b. Update bandwidth and frequency parameters

[0023] For the bandwidth parameter μ k :

[0024]

[0025] For the frequency parameter ω k :

[0026]

[0027] Among them, m t and v t are the adaptive estimates of the first-order and second-order moments, α is the learning rate for bandwidth update, and α ω is the learning rate of the frequency update, which are all adaptive adjustment parameters;

[0028] Adaptive learning rate adjustment is added to the improved Adam optimization algorithm so that the learning rate is automatically adjusted according to the gradient change. The calculation formula for adaptive learning rate adjustment is:

[0029]

[0030] Among them, δ is the adjustment factor, which is 10 -6 ,|g t | is the absolute value of the current gradient. After introducing the adaptive learning rate adjustment, the learning rate is appropriately reduced when the gradient is large to avoid drastic updates, and appropriately increased when the gradient is small to speed up the convergence;

[0031] Using the improved Adam optimization algorithm and adaptive learning rate adjustment, u k (t) and ω k Perform iterative updates to minimize the objective function. In each round of updates, check whether the bandwidth change meets the preset convergence conditions. If it does, the iteration ends. Finally, after multiple iterations, K converged intrinsic mode functions are obtained, which are expressed as:

[0032]

[0033] Furthermore, the specific method of using the Lasso regression model in step 2 to screen out the IMFs with higher importance as information carriers is:

[0034] The decomposed IMFs are used to form a feature matrix X, where each column represents an IMF and each row represents a time node. The corresponding target variable y, i.e., the label of the fault state, is set;

[0035] The goal of Lasso regression is to minimize the following loss function:

[0036]

[0037] Where n is the number of samples, y i is the target variable, X ij is the element of the feature matrix, β 0 is the intercept, β j is the regression coefficient of the IMF on the target variable, and λ is the regularization parameter, which controls the constraint strength on the regression coefficient. The larger it is, the stronger the constraint is.

[0038] The regularization parameter λ is used to balance the deviation and variance of the model, and the training data is used to fit the Lasso regression model to obtain the regression coefficient β of the IMF. j , the training data is the IMF feature matrix X and the target variable y. By adjusting the regularization parameter, the number of IMFs to be screened is controlled; the regression coefficient β of each IMF is compared. jThe importance of IMF in fault diagnosis is evaluated, a threshold is set, and IMF with large non-zero coefficients is screened out as information carriers for fault diagnosis of photovoltaic systems.

[0039] Furthermore, in step 3, the selected IMF is subjected to wavelet transformation and decomposed into low-frequency components and high-frequency components in different frequency bands. The specific method is as follows:

[0040] The selected IMF components are decomposed using discrete wavelet transform DWT. Each level of wavelet decomposition includes two filtering steps, namely low-pass filtering and high-pass filtering. In the j-th level decomposition, the low-frequency coefficients are expressed as:

[0041]

[0042] The high frequency coefficients are expressed as:

[0043]

[0044] Among them, A j represents the low-frequency component of the jth layer, D j represents the high-frequency component of the jth layer, h and g represent the low-pass and high-pass filter coefficients of the wavelet, respectively. After J-level wavelet decomposition, the multi-scale subsequence A is obtained. J and D1,D2,…,D J Among them, A J It is the lowest frequency approximate signal, indicating the main trend of the signal and suitable for further coarse-grained processing; D J is the detail signal of the jth layer, representing the high-frequency component of the layer, reflecting the local fluctuation and mutation information in the current signal, and can identify short-term changes and noise; when decomposed to the Lth layer, the original signal is decomposed into the low-frequency coefficients of the Lth layer and the high-frequency coefficients of multiple scales, expressed as:

[0045]

[0046] Among them, A L is the low-frequency component when decomposed to the Lth layer, D l It is the high-frequency component of the lth layer when decomposed to the Lth layer.

[0047] Furthermore, the low-frequency components obtained after wavelet transform are coarse-grained, and then the discrete entropy values ​​of the coarse-grained subsequences of different time scales are calculated. The specific method is as follows:

[0048] First, the time scale factor is determined by considering the different time scales of different phenomena involved in the output current signal, and the low-frequency components obtained after wavelet transform are coarse-grained according to the proportion of the scale factor; then the low-frequency components obtained after wavelet transform are grouped according to the proportion of the time scale factor, and the data in the group are averaged, and finally a new coarse-grained time series is obtained, for each length of N / τ cs Create a new time series with the scale:

[0049]

[0050] Among them, τ sc is the time scale factor, which determines how many subsequences the low-frequency component is coarse-grained into. For each time scale factor τ sc , group the data in the low-frequency component, average each group of data, and get the new data point y j (τsc) , which represents the original time series at τ sc As a result of the coarse-graining on the time scale, the length of the coarse-grained time series decreases as the time scale ratio increases.

[0051] Furthermore, the coarse-grained subsequences of different time scales are decomposed to perform multi-scale discrete entropy MDE calculation. The discrete entropy MDE is defined as follows:

[0052]

[0053] Among them, p i represents the probability of the system in unit i, and n represents the characteristic dimension of the signal.

[0054] Furthermore, in step 5, the trend change of the discrete entropy value curve at different time scales is observed, and the health of the photovoltaic system is judged according to the shape and threshold of the curve. The specific method for diagnosing whether the photovoltaic system has a shading fault is as follows:

[0055] First, observe the shape of the curve to preliminarily determine whether there is a fault in the photovoltaic system. When the photovoltaic system is in clear weather, the MDE curve is usually bell-shaped; when the photovoltaic system is in cloudy weather, the MDE curve is usually flat; when there is a fault in the photovoltaic system, the MDE curve will shift from the normal bell-shaped curve and move up or down; when a fault is observed in the photovoltaic system, compare the discrete entropy curves of fault-free sunny days, fault-free cloudy days and faulty sunny days, and obtain the MDE value of the faulty sunny day that is significantly different from the fault-free sunny day and the fault-free cloudy day, and set it as the threshold. The threshold is used to determine whether the photovoltaic system has a shading fault and whether the shading fault is starting or ending, thereby achieving the effect of fault diagnosis.

[0056] Beneficial effects:

[0057] (1) The present invention overcomes the limitations of the adaptive variational mode decomposition algorithm through an improved adaptive variational mode decomposition algorithm. The improved AVMD overcomes the limitations of AVMD setting fixed learning rate, bandwidth and frequency parameter initial values, introduces adaptive learning rate adjustment, and optimizes and updates the bandwidth parameter and frequency parameter in combination with the improved Adam algorithm, making the decomposition process more flexible and better adapted to the characteristics of the signal. It has higher computational efficiency and convergence speed when processing non-stationary and nonlinear signals, and can more accurately decompose the current signal into intrinsic mode functions. And using the advantages of entropy, multi-scale entropy is used in the field of diagnosis. The multi-scale method can evaluate the entropy of different phenomena at different time scales. The above technology is used to detect faults on photovoltaic equipment. The AVMD algorithm that introduces the Adam optimization algorithm and the adaptive learning rate is used to extract the current signal carrying the photovoltaic system status and fault characteristic information, and then the photovoltaic system faults at different time scales are diagnosed by multi-scale discrete entropy. The method has good robustness and is easy to implement.

[0058] (2) The present invention utilizes the current data collected by the inverter and adopts an adaptive variational mode decomposition algorithm that introduces the Adam optimization algorithm and an adaptive learning rate to decompose the current signal into eigenmode functions (IMFs). The algorithm has high accuracy, is independent of weather changes, and has good robustness.

[0059] (3) The present invention performs coarse-graining processing on the low-frequency components obtained after wavelet transform. Coarse-graining processing is a signal processing technology that creates a new time series to ensure that the length of each time series meets the conditions for calculating discrete entropy, thereby improving the reliability of the entropy value and the efficiency of the calculation.

[0060] (4) MDE can analyze the signal complexity at different time scales and evaluate the health status of the photovoltaic system more comprehensively. It can reduce the impact of noise by coarse-graining the signal. Compared with other entropy calculation methods, discrete entropy has higher calculation efficiency and is suitable for real-time monitoring.

[0061] (5) The improved adaptive variational mode decomposition algorithm and the multi-scale discrete entropy algorithm are combined to diagnose photovoltaic system faults and distinguish the characteristics of different types of faults at different time scales, which can improve the accuracy and efficiency of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is an example diagram of a process based on an improved adaptive variational mode decomposition algorithm and a combination of multi-scale discrete entropy in the present invention;

[0063] Figure 2It is the time scale factor and normalized discrete entropy value curve of the current signal output by the photovoltaic array under normal and fault conditions after being processed by the improved AVMD-MDE in the present invention;

[0064] Figure 3 is an example diagram of a process of adaptive variational modal decomposition based on improvement of current signal in the present invention;

[0065] Figure 4 It is an example diagram of the process of calculating discrete entropy in the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The photovoltaic array fault diagnosis method based on the current signal output of the photovoltaic power station disclosed in the present invention comprises the following steps:

[0068] Step 1: Collect the current signal transmitted by the inverter on the photovoltaic power station. The inverter converts the direct current generated by the photovoltaic array into alternating current for access to the grid. The photovoltaic power station consists of two groups of 12 modules, which are composed of 72 solar cells, each group of 18 cells is connected in series and equipped with a bypass diode. The solar panels are tilted 26° from the ground and face north. Two groups of lines are tested. Line 1 is used to collect current data for shadow faults, and the other line is used to collect data for normal weather. By comparing the difference between the two groups of data curves, the fault curve graph is obtained. The calculated power of line 1 is 2.04KWp, and the calculated power of line 2 is 2.09KWp. The characteristics of the photovoltaic modules are shown in Table 1. In order to implement the conversion and security of data transmission, a data pipeline consisting of a gateway, a network service, a database for storing data, and an application server is implemented. The electrical data measurement is obtained by the operation and management system of the photovoltaic power station connected to the grid and coupled with the inverter. The frequency of data transmission is selected by the telemetry control of the software. All radiation and meteorological sensors are connected to the data logger for data collection, and the data is then transmitted to the server via the TCP / IP protocol. In order to be able to use the data in real time, the data is stored at a sampling frequency of 1HZ. In order to ensure the real-time availability of data, the software is written in C language, monitoring data files from various sources and injecting the data into the database in the form of time series. Therefore, an end-to-end data pipeline relying on open source technology was developed, and data from sensors and inverters are regularly sent, collected and stored in a centralized database system.

[0069] Table 1

[0070]

[0071] Step 2: Use Kalman filtering to denoise the current signal output by the inverter. Use the improved adaptive variational mode decomposition algorithm (improved AVMD algorithm), that is, use the improved Adam optimization algorithm to replace the traditional optimization algorithm, decompose the denoised current signal into multiple intrinsic mode functions (IMFs), and iteratively update the bandwidth parameter during the decomposition process. The specific steps include:

[0072] The inverter connected to the grid through the operation and management system of the photovoltaic power station uses the current data collected by the inverter, which contains the current signal output by the photovoltaic module blocked by the shadow. The current data output by the inverter is Kalman denoised to generate the denoised current signal. The denoised current signal is decomposed using the improved AVMD algorithm. The improved AVMD algorithm is an adaptive variational mode decomposition algorithm for signal processing, which can decompose complex signals into several intrinsic mode functions. The improved AVMD overcomes the limitations of AVMD in setting fixed initial values ​​of learning rate, bandwidth and frequency parameters, introduces adaptive learning rate adjustment, and combines the improved Adam algorithm to optimize and update the bandwidth parameters and frequency parameters, making the decomposition process more flexible.

[0073] Kalman filtering requires setting the initial state and the initial value of the noise covariance matrix, defining the process noise covariance and the measurement noise covariance, which are used to control the filtering strength and signal smoothness. Construct a state space model of the Kalman filter, predict the state and error covariance of the next moment based on the current state and noise model, and then use the measured value at the current moment to update the predicted state and reduce the prediction error. In the denoising stage, the current signal is continuously predicted and updated, and the Kalman filter denoising is continuously iterated to generate a denoised current signal sequence.

[0074] For the current signal after Kalman filtering, the goal of improving AVMD is to decompose it into K modal functions u k (t) and the corresponding center frequency ω k , the decomposition objective is to minimize the following bandwidth-constrained variational formula:

[0075]

[0076] in, represents the time derivative, which is used to represent the change of bandwidth, u k (t) is the Kth mode function of the signal, ω k is the center frequency of each mode function, which is dynamically updated with the adaptive learning rate and improved Adam optimization.

[0077] The Lagrangian multiplier λ(t) is introduced in the improved AVMD, so that the sum of the modal components can reconstruct the original signal, making the reconstructed signal as close to the original signal as possible to ensure the integrity of the signal. The Lagrangian function is defined as:

[0078]

[0079] Among them, α is the adaptive learning rate, which is used to dynamically control the optimization step size of the bandwidth, λ(t) is the Lagrange multiplier, which ensures the reconstruction accuracy of the signal, f(t) is the input original current signal of the photovoltaic power station, and K is the number of modal components, which represents the number of intrinsic mode functions obtained when decomposing the signal.

[0080] In order to solve the above optimization problem, an improved Adam optimization algorithm is introduced to optimize u k (t) and ω k First, the Adam optimization algorithm needs to initialize parameters, set the initial learning rate and two exponential decay rate parameters β 1 and β 2 , initialize the first-order moment (momentum) m 0 = 0 and the second moment (weighted non-central variance) v 0 =0, set a small value ∈ to avoid division by zero error.

[0081] The core calculation steps of the improved Adam optimization algorithm are as follows:

[0082] a. Calculate the gradient

[0083]

[0084] b. Update bandwidth and frequency parameters

[0085] For the bandwidth parameter μ k :

[0086]

[0087] For the frequency parameter ω k :

[0088]

[0089] Among them, m t and v t are the adaptive estimates of the first-order and second-order moments, α is the learning rate for bandwidth update, and α ω is the learning rate for frequency update, both of which are adaptive adjustment parameters.

[0090] Adaptive learning rate adjustment is added to the improved Adam optimization algorithm so that the learning rate is automatically adjusted according to the gradient change. The calculation formula for adaptive learning rate adjustment is:

[0091]

[0092] Where δ is the adjustment factor, usually 10 -6 ,|g t | is the absolute value of the current gradient. After introducing adaptive learning rate adjustment, the learning rate can be appropriately reduced when the gradient is large to avoid drastic updates. At the same time, it can also be appropriately increased when the gradient is small to speed up convergence.

[0093] Using the improved Adam optimization algorithm and adaptive learning rate adjustment, u k (t) and ω k Perform iterative updates to minimize the objective function. In each round of updates, check whether the bandwidth change meets the preset convergence conditions. If it does, the iteration ends. Finally, after multiple iterations, K converged intrinsic mode functions are obtained, which are expressed as:

[0094]

[0095] The optimization process of the improved adaptive variational mode decomposition mainly consists of the following steps:

[0096] Step a: Initialize the modal function of each mode at iteration number 1 and the center frequency ω k and the Lagrange multiplier Spectra and center frequencies of each K modes.

[0097] Step b: In the improved Adam optimization algorithm, the mode function is updated using momentum:

[0098]

[0099] Step c: Update the Lagrange multiplier:

[0100]

[0101] Step d: Update the center frequency

[0102]

[0103] Step e: Repeat steps b to e to determine whether the bandwidth change meets the convergence condition. If the convergence condition is met, stop the iteration.

[0104] The improved AVMD algorithm no longer directly relies on the alternating direction multiplier method to update the frequency and bandwidth parameters, but uses the improved Adam optimization algorithm and adaptive learning rate to gradually adjust u k (t) and center frequency ω kThe value of . First, initialize the center frequency, modal function and Lagrange multiplier of each mode. Then, in each iteration, update the modal function, Lagrange multiplier and center frequency in turn. The modal function is updated by the improved Adam optimization algorithm. The Lagrange multiplier is updated to maintain the bandwidth constraint. The center frequency is updated to ensure that the frequency can accurately reflect the main components of the signal. Finally, it is determined whether the change in bandwidth meets the convergence condition. If converged, stop the iteration; otherwise, return to the iteration step to continue the iteration. Compared with the original adaptive variational modal decomposition algorithm, the adaptive variational modal decomposition method with the improved Adam optimization algorithm and adaptive learning rate adjustment can adjust the modal parameters and frequency parameters more quickly and stably in each iteration, and decompose the signal into several intrinsic mode functions with specific bandwidth and center frequency.

[0105] The improved AVMD algorithm needs to initialize some parameters at the beginning, set the initial mode function to 0, and the initial center frequency to 0.06. In the iterative process, the improved Adam optimization algorithm and adaptive learning rate are used to gradually adjust u k (t) and center frequency ω k The value of . In each iteration, the modal function, Lagrange multiplier and center frequency are updated in turn. The modal function is updated by the improved Adam optimization algorithm. The Lagrange multiplier is updated to maintain the bandwidth constraint. The center frequency is updated to ensure that the frequency can accurately reflect the main components of the signal. Finally, it is determined whether the change in bandwidth meets the convergence condition. If converged, the iteration is stopped; otherwise, the iteration step is returned to continue the iteration. The AVMD algorithm that introduces the improved Adam optimization algorithm and adaptive learning rate is more efficient than the original AVMD algorithm in processing non-stationary and multi-modal signals, and has more advantages in practical applications.

[0106] Step 3: Decompose the current signal into multiple intrinsic mode functions (IMFs), and select the IMFs with higher importance as information carriers according to the Lasso regression model. The specific steps include:

[0107] The IMF decomposed by the improved AVMD algorithm is processed and the decomposed IMF is formed into a feature matrix X. Each column represents an IMF and each row represents a time node. The corresponding target variable y, that is, the label of the fault state, is set.

[0108] Lasso regression is a linear regression model that can effectively perform feature selection after introducing L1 regularization. The goal of Lasso regression is to minimize the following loss function:

[0109]

[0110] Where n is the number of samples, yi is the target variable, X ij is the element of the feature matrix, β 0 is the intercept, β j is the regression coefficient of the IMF on the target variable, and λ is the regularization parameter, which controls the constraint strength on the regression coefficient. The larger it is, the stronger the constraint is.

[0111] The regularization parameter λ is used to balance the deviation and variance of the model, and the training data (IMF feature matrix and target variable) are used to fit the Lasso regression model to obtain the IMF regression coefficient β j By adjusting the regularization parameter, the number of IMFs to be screened can be controlled. Comparing the regression coefficient β of each IMF j The importance of IMF in fault diagnosis is evaluated by the size of IMF. Generally, a larger absolute value indicates that IMF has a greater impact on the target variable. IMFs with more information are selected as information carriers for diagnosis, and thresholds are set to select IMFs with larger non-zero coefficients as information carriers for fault diagnosis of photovoltaic systems.

[0112] Step 4: Perform wavelet transform on the selected IMF to reduce the complexity of subsequent calculation of discrete entropy. Wavelet transform decomposes the original time series of the output current of the photovoltaic power station over time into low-frequency components and high-frequency components in different frequency bands.

[0113] Wavelet transform is a time-frequency localization analysis method that can decompose the signal into different time and frequency components. The selected IMF components are decomposed using discrete wavelet transform (DWT). The DWT decomposition process is obtained through cascaded filtering operations. Each level of wavelet decomposition includes two filtering steps, namely low-pass filtering and high-pass filtering. In the j-th level of decomposition, the low-frequency coefficient can be expressed as:

[0114]

[0115] The high frequency coefficient can be expressed as:

[0116]

[0117] Among them, A j represents the low-frequency component of the jth layer, D j represents the high-frequency component of the jth layer, h and g represent the low-pass and high-pass filter coefficients of the wavelet, respectively. After J-level wavelet decomposition, the multi-scale subsequence A is obtained J and D1,D2,…,D J Among them, A J It is the lowest frequency approximate signal, indicating the main trend of the signal and suitable for further coarse-grained processing. Jis the detail signal of the jth layer, representing the high-frequency component of the layer, reflecting the local fluctuation and mutation information in the current signal, and can identify short-term changes and noise. These subsequences provide multi-scale information from high frequency to low frequency, which helps to capture the changing characteristics of the signal at different time scales. When decomposed to the Lth layer, the original signal is decomposed into low-frequency coefficients of the Lth layer and high-frequency coefficients of multiple scales, expressed as:

[0118]

[0119] Step 5: In order to further capture the changing characteristics of each subsequence at different time scales, the low-frequency components obtained after wavelet transform are coarse-grained, and then multi-scale discrete entropy (MDE) is calculated based on the coarse-grained subsequences of different time scales decomposed. Considering the different time scales involved in the current signal, the discrete entropy of the coarse-grained subsequences of different time scales is calculated to obtain the discrete entropy value, and then the calculated discrete entropy value is normalized. Its parameter selection and results are as follows:

[0120] The multi-scale method is a very common signal analysis method, which can be used to observe and analyze signal characteristics at different time scales. In this patent, the multi-scale method takes into account the different time scales of different phenomena involved in the output current signal, and performs coarse-graining processing on the time scale for the low-frequency components obtained after wavelet transformation. cs Create a new time series with the same scale.

[0121]

[0122] Among them, τ sc is the time scale factor, which determines how many subsequences the low-frequency component is coarse-grained into. For each time scale factor τ sc , group the data in the low-frequency component, average each group of data, and get the new data point y j (τsc) , which represents the original time series at τ sc Coarse-grained results on the time scale.

[0123] Multiscale discrete entropy is a multiscale method based on discrete entropy, taking into account the different time scales of different phenomena involved in the output signal. Entropy is a statistical algorithm that quantifies the probability of adjacent points in a time series being within a predetermined range, and can thus measure the complexity and information richness of a time series. Entropy is defined as follows:

[0124]

[0125] Among them, p iis the probability that the system is located in unit i, and n represents the characteristic dimension of the signal.

[0126] The discrete entropy algorithm is the latest algorithm developed, and the flowchart is as follows: Figure 4 As shown, the specific steps of the algorithm are as follows:

[0127] Step 1: Class Mapping In this work, the original signal is mapped to the Normal Cumulative Distribution Function (NCDF): x = {x 1 ,x 2 ,…,x N}→y={y 1 ,y 2 ,...,y N}, the values ​​of y range from 0 to 1. In order to assign an integer value from 1 to c to each value of yi, an algorithm is used, where c is the number of categories.

[0128]

[0129] Step 2: Create embedding vector and dispersion pattern The created embedding vector has a length of m and takes into account the time delay d.

[0130]

[0131] Mapping to dispersion modes in The number of possible patterns is equal to c m .

[0132] Step 3: Relative frequency calculation The relative frequency of each pattern is defined as the number of times each pattern appears divided by the total number of embedding vectors:

[0133]

[0134] Step 4: Entropy calculation based on entropy definition

[0135]

[0136] In the present invention, in order to facilitate numerical comparison, the discrete entropy value is normalized by the following formula:

[0137]

[0138] The parameters for DE need to be determined and the following recommendations must be considered:

[0139] (1) d should be set to 1 to avoid aliasing effects;

[0140] (2)c m The length must be less than the length of the time series;

[0141] (3) c must be greater than 1 to have different modes. If c is too small, very different values ​​may be assigned to the same mode; if c is too large, DE will be sensitive to noise;

[0142] (4) c can be selected between 4 and 8;

[0143] (5) m must be carefully chosen; when the value of m is too small, no kinetic changes can be detected; when the value of m is too large, no small kinetic changes can be detected;

[0144] (6) The larger the values ​​of m and c, the longer the calculation time;

[0145] Considering the multi-scale method, the original time series of the output current of the photovoltaic power station over time is coarse-grained, and each time series of length N / τ is cs Create a new time series with the same scale.

[0146] Then, the discrete entropy of each coarse-grained time series is calculated. The length of the coarse-grained time series decreases as the time scale ratio increases. According to relevant literature, the length of the time series must be within 10 m To 20 m Only then can a reliable entropy value be obtained.

[0147] The time delay d, the number of categories c, and the length of the embedding vector are chosen according to previous suggestions. The maximum scaling factor τ s The definitions of c and the time series length N are the result of the following considerations:

[0148] 1. The size of the time series should be short enough to minimize the necessary data volume and computing time and maximize monitoring efficiency. The window width should not exceed 30 minutes at most.

[0149] 2. The frequency of the photovoltaic power plant acquisition system studied is very low, and no external measurement equipment is added. In 30 minutes, about 300 data points are collected.

[0150] 3. Dispersion entropy requires a minimum number of points to ensure that the entropy value is valid. Considering the length of the coarse-grained time series, N / τ s c must be significantly higher than c m =6 2 =36.

[0151] Table 4 lists all the parameters of multiscale dispersion entropy.

[0152]

[0153] Table 4

[0154] Step 6: Calculate the multi-scale discrete entropy of each IMF based on the data in Table 4, select the IMF graph with the largest amount of information, and analyze it:

[0155] Figure 2 a is the experimental data set. The results of the fault-free sunny data are represented by the black dotted line, the fault-free cloudy data are represented by the black solid line, and the fault data are represented by the black dotted line. The curve with a triangle represents the fault occurrence point. The downward curve represents the fault starting point, and the upward curve represents the fault end point. Figure 2 As shown in (b), the IMF5 MDE curves show three distinct patterns according to different experimental conditions. The sunny day fault data and the no-fault data produce a bell-shaped curve, while the fault data produce a shifted curve of the sunny day data. In fact, Figure 2 a shows that the output current changes greatly in the time period around ten o'clock, which also explains the results of the improved AVMD-MDE. In order to evaluate the performance of the improved AVMD-MDE in terms of shading faults, two types of data were collected, one for cloudy days and the other for sunny days, and analyzed using the improved AVMD-MDE algorithm.

[0156] Figure 2 c shows two time periods: cloudy and sunny. The time period between 8:00 and 10:00 is sunny, and the time period after that is cloudy. Figure 2 d shows the results of the MDE analysis, showing two different MDE curve patterns, a straight line corresponding to cloudy days and a bell-shaped curve corresponding to sunny days. One of the curves is incorrectly clustered, corresponding to the tenth time period. In fact, this time period corresponds to a continuous low irradiance, which leads to low current generation and low change, thus explaining the trend of the curve. Figure 2 e describes the sunny part of the day, the occurrence of the fault, and the cloudy day after the fault. Figure 2 f shows that the sunny and cloudy curves have different trends. In the case of a fault, it can be observed that the MDE curve is transformed from the healthy sunny curve, such as Figure 2 However, compared with the cloudy sky curve, the fault start and end curves have the same trends and values.

[0157] As shown in Table 5, a total of 36 data sets were collected. These preliminary results show the relationship between the MDE value and the system health. Based on the curves shown previously, two thresholds can be retained to accurately define the PV system health. Figure 2f It can be seen that when the scale factor is 1, the DE value below the critical value of 0.6 indicates a cloudy sky or the beginning or end of a fault. Although the DE values ​​on sunny days are very close for both faulty and non-faulty data at scale factor 1, the DE values ​​are significantly different at scale factors 4 and 5. In fact, at scale factor 4, a DE value above 0.9 indicates a faulty state. In this work, scale factors 1 and 4 are selected with thresholds of 0.6 and 0.9, respectively, to determine the health status of the photovoltaic system. After evaluation, the overall accuracy of the proposed method is 94.30%. In this invention, the health status is determined by a simple rule-based diagnosis, and the selected scale and related thresholds are set based on the observation results.

[0158] Table 5

[0159]

[0160]

[0161] In summary, the present invention has developed an innovative online fault detection and diagnosis method. The method relies on an improved adaptive variational mode decomposition algorithm (improved AVMD) and multi-scale discrete entropy (MDE). The entropy value of a specific scale is selected as a quantitative indicator to determine the health status of the system. The diagnosis based on the improved AVMD-MDE does not require additional equipment and can be easily implemented online. Combining the improved adaptive variational mode decomposition and multi-scale discrete entropy has broad prospects and can be applied to more complex fault diagnosis in the future. The flow chart based on the improved adaptive variational mode decomposition and multi-scale discrete entropy is shown in the figure. Figure 1 shown.

[0162] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic array fault diagnosis method based on current signal output of a photovoltaic power station, characterized in that: The following steps are involved: Step 1. Connect the operation and management system of the photovoltaic power station to the inverter of the power grid to collect current signals, collect current signals of photovoltaic modules blocked by shadows and current signals in normal weather, use Kalman filtering to denoise the input current signals, and use the improved AVMD algorithm to decompose the denoised current signals into multiple intrinsic mode functions IMFs; the improved AVMD algorithm introduces adaptive learning rate adjustment, and combines the improved Adam algorithm to optimize and update the bandwidth parameters and frequency parameters; Step 2. Apply the Lasso regression model to select the best regularization parameter using cross validation to balance the bias and variance of the model, fit the Lasso regression model using the training data, calculate the regression coefficient of the IMF, evaluate the importance of the IMF in fault diagnosis, and then determine the IMFs with higher importance as the information carrier for diagnosing photovoltaic system faults; Step 3. Perform wavelet transform on the selected IMF to decompose the original time series of the output current of the photovoltaic power station into low-frequency components and high-frequency components in different frequency bands; Step 4. Coarse-graining the low-frequency components, and then calculating the multi-scale discrete entropy MDE on the coarse-grained subsequences of different time scales, and normalizing the calculated discrete entropy values; Step 5. Compare the normalized discrete entropy value curves of fault-free sunny days, fault-free cloudy days, and fault-free sunny days, observe the differences in discrete entropy values ​​at different time scales, find out the discrete entropy value with a large difference between the photovoltaic system with shading fault and the one without fault, set it as the threshold, and use it to judge whether the photovoltaic system is in a normal state, a shading fault state, and whether the shading fault occurs or ends, so as to realize fault diagnosis of the photovoltaic system.

2. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 1, characterized in that: The specific method of decomposing the current signal output by the photovoltaic power station inverter into multiple intrinsic mode functions using the improved AVMD in step 1 is: For the current signal after Kalman filtering, the goal of improving AVMD is to decompose the current signal into K modal functions u k (t) and the corresponding center frequency ω k , the decomposition objective is to minimize the following bandwidth-constrained variational formula: in, represents the time derivative, which is used to represent the change of bandwidth, u k (t) is the kth mode function of the signal, ω k is the center frequency of each mode function, which is dynamically updated with adaptive learning rate and improved Adam optimization; The Lagrangian multiplier λ(t) is introduced in the improved AVMD, so that the sum of the modal components can reconstruct the original signal, making the reconstructed signal as close to the original signal as possible to ensure the integrity of the signal. The Lagrangian function is defined as: Among them, α is the adaptive learning rate, which is used to dynamically control the optimization step size of the bandwidth, λ(t) is the Lagrange multiplier, which ensures the reconstruction accuracy of the signal, f(t) is the input original current signal of the photovoltaic power station, and K is the number of modal components, which represents the number of intrinsic mode functions obtained when decomposing the signal; Introduce the improved Adam optimization algorithm to optimize u k (t) and ω k Dynamic update, first, the Adam optimization algorithm initializes the parameters, sets the initial learning rate and two exponential decay rate parameters β1 and β2, initializes the first-order moment m0 ​​= 0 and the second-order moment v0 = 0, and sets a small value ∈ to avoid division by zero errors; The calculation steps of the improved Adam optimization algorithm are as follows: a. Calculate the gradient b. Update bandwidth and frequency parameters For the bandwidth parameter μ k : For the frequency parameter ω k : Among them, m t and v t are the adaptive estimates of the first-order and second-order moments, α is the learning rate for bandwidth update, and α ω is the learning rate of the frequency update, which are all adaptive adjustment parameters; Adaptive learning rate adjustment is added to the improved Adam optimization algorithm so that the learning rate is automatically adjusted according to the gradient change. The calculation formula for adaptive learning rate adjustment is: Among them, δ is the adjustment factor, which is 10 -6 ,|g t | is the absolute value of the current gradient. After introducing the adaptive learning rate adjustment, the learning rate is appropriately reduced when the gradient is large to avoid drastic updates, and appropriately increased when the gradient is small to speed up the convergence; Using the improved Adam optimization algorithm and adaptive learning rate adjustment, u k (t) and ω k Perform iterative updates to minimize the objective function. In each round of updates, check whether the bandwidth change meets the preset convergence conditions. If it does, the iteration ends. Finally, after multiple iterations, K converged intrinsic mode functions are obtained, which are expressed as:

3. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 1, characterized in that: The specific method of using the Lasso regression model in step 2 to select the IMF with higher importance as the information carrier is: The decomposed IMFs are used to form a feature matrix X, where each column represents an IMF and each row represents a time node. The corresponding target variable y, i.e., the label of the fault state, is set; The goal of Lasso regression is to minimize the following loss function: Where n is the number of samples, y i is the target variable, X ij is the element of the feature matrix, β0 is the intercept, β j is the regression coefficient of the IMF on the target variable, and λ is the regularization parameter, which controls the constraint strength on the regression coefficient. The larger it is, the stronger the constraint is. The regularization parameter λ is used to balance the deviation and variance of the model, and the training data is used to fit the Lasso regression model to obtain the regression coefficient β of the IMF. j , the training data is the IMF feature matrix X and the target variable y. By adjusting the regularization parameter, the number of IMFs to be screened is controlled; the regression coefficient β of each IMF is compared. j The importance of IMF in fault diagnosis is evaluated, a threshold is set, and IMF with large non-zero coefficients is screened out as information carriers for fault diagnosis of photovoltaic systems.

4. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 1, characterized in that: In step 3, the selected IMF is decomposed into low-frequency components and high-frequency components in different frequency bands by wavelet transformation. The specific method is as follows: The selected IMF components are decomposed using discrete wavelet transform DWT. Each level of wavelet decomposition includes two filtering steps, namely low-pass filtering and high-pass filtering. In the j-th level decomposition, the low-frequency coefficients are expressed as: The high frequency coefficients are expressed as: Among them, A j represents the low-frequency component of the jth layer, D j represents the high-frequency component of the jth layer, h and g represent the low-pass and high-pass filter coefficients of the wavelet, respectively. After J-level wavelet decomposition, the multi-scale subsequence A is obtained. J and D1,D2,…,D J Among them, A J It is the lowest frequency approximate signal, indicating the main trend of the signal and suitable for further coarse-grained processing; D J is the detail signal of the jth layer, representing the high-frequency component of the layer, reflecting the local fluctuation and mutation information in the current signal, and can identify short-term changes and noise; when decomposed to the Lth layer, the original signal is decomposed into the low-frequency coefficients of the Lth layer and the high-frequency coefficients of multiple scales, expressed as: Among them, A L is the low-frequency component when decomposed to the Lth layer, D l It is the high-frequency component of the lth layer when decomposed to the Lth layer.

5. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 1, characterized in that: The low-frequency components obtained after wavelet transform are coarse-grained, and then the discrete entropy values ​​of the coarse-grained subsequences of different time scales are calculated. The specific method is as follows: First, the time scale factor is determined by considering the different time scales of different phenomena involved in the output current signal, and the low-frequency components obtained after wavelet transform are coarse-grained according to the proportion of the scale factor; then the low-frequency components obtained after wavelet transform are grouped according to the proportion of the time scale factor, and the data in the group are averaged, and finally a new coarse-grained time series is obtained, for each length of N / τ cs Create a new time series with the scale: Among them, τ sc is the time scale factor, which determines how many subsequences the low-frequency component is coarse-grained into. For each time scale factor τ sc , group the data in the low-frequency component, average each group of data, and get the new data point y j (τsc) , which represents the original time series at τ sc As a result of the coarse-graining on the time scale, the length of the coarse-grained time series decreases as the time scale ratio increases.

6. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 5, characterized in that: The coarse-grained subsequences of different time scales are decomposed to perform multi-scale discrete entropy MDE calculation. The discrete entropy MDE is defined as follows: Among them, p i represents the probability of the system in unit i, and n represents the characteristic dimension of the signal.

7. A photovoltaic array fault diagnosis method based on photovoltaic power station current signal output according to claim 1, characterized in that: In step 5, the trend change of the discrete entropy value curve at different time scales is observed, and the health of the photovoltaic system is judged according to the shape and threshold of the curve. The specific method for diagnosing whether the photovoltaic system has a shading fault is as follows: First, observe the shape of the curve to preliminarily determine whether there is a fault in the photovoltaic system. When the photovoltaic system is in clear weather, the MDE curve is usually bell-shaped; when the photovoltaic system is in cloudy weather, the MDE curve is usually flat; when there is a fault in the photovoltaic system, the MDE curve will shift from the normal bell-shaped curve and move up or down; when a fault is observed in the photovoltaic system, compare the discrete entropy curves of fault-free sunny days, fault-free cloudy days and faulty sunny days, and obtain the MDE value of the faulty sunny day that is significantly different from the fault-free sunny day and the fault-free cloudy day, and set it as the threshold. The threshold is used to determine whether the photovoltaic system has a shading fault and whether the shading fault is starting or ending, thereby achieving the effect of fault diagnosis.

Images