Islanding detection method for photovoltaic power station based on wavelet transform and GRU neural network

Abstract

The application belongs to the technical field of photovoltaic power station systems, and discloses a photovoltaic power station island detection method based on wavelet transform and a GRU neural network, which comprises the following steps: firstly, adopting wavelet packet decomposition technology to process the original voltage signal waveform collected; secondly, inputting the voltage signal filtered through WPT into a variational mode decomposition module, and adaptively decomposing the filtered voltage signal into multiple mode components with different center frequencies through a VMD algorithm; finally, inputting the relevant data of the mode components obtained through decomposition into a GRU neural network model, and completing the learning training and classification judgment of the voltage signal feature data, so as to finally realize the accurate detection and identification of the island phenomenon. The photovoltaic power station island detection method based on wavelet transform and the GRU neural network combines wavelet packet decomposition, variational mode decomposition and the GRU neural network organically, forms a set of efficient and accurate island detection method, and can realize the accurate identification of the island working condition, thereby providing protection for the stable operation of the power grid.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photovoltaic power station systems, and relates to a photovoltaic power station island detection method based on wavelet transform and a GRU neural network. BACKGROUND

[0002] In recent years, with the continuous expansion of the construction scale of photovoltaic power stations, the capacity of new energy power generation systems integrated into the power grid has gradually increased, which has brought many challenges to the stable operation of traditional power systems. Among them, the unplanned island phenomenon has become a prominent problem that needs to be solved urgently. Unplanned island refers to the non-normal operating state of a distributed photovoltaic power station that is disconnected from the main power grid without following the planned scheme, but still continuously supplies power to local power loads. The existence of this phenomenon poses a serious threat to the stable operation and safe operation of the power grid, and may cause potential risks such as equipment damage and power supply interruption.

[0003] At present, the commonly used island detection methods in the industry are mainly divided into two categories: active detection method and passive detection method. Within the scope of the active detection method, specific technical means such as reactive power disturbance method and Sandia frequency shift method (SFS) are included. The advantage of this method is that the detection blind area is relatively small, and it can cover most of the possible island scenarios. However, its disadvantages are that it will have an adverse effect on power quality, the cost is high during application, and when multiple inverters are in parallel operation, it may also cause detection failure and fail to accurately identify the island working condition.

[0004] Compared with the active detection method, the passive detection method does not need to apply additional disturbance to the power grid, so it will not affect the power quality and has good operation compatibility. However, this method has a large detection blind area, which is easy to miss some island scenarios, and the threshold parameters required during detection are difficult to accurately determine, which to some extent reduces the reliability of the detection results.

[0005] Based on the characteristics and shortcomings of the above existing technologies, to minimize the disturbance to power quality, the inventors of the present application prefer to use the passive island detection method in the selection of detection methods. At the same time, in order to further improve the accuracy of island detection and solve the inherent defects of the passive detection method, an intelligent passive island detection algorithm combining signal processing technology and intelligent algorithm is planned to be introduced, and a new detection scheme is proposed. SUMMARY

[0006] The application aims to overcome the defects of the prior art and provide a photovoltaic power station island detection method based on wavelet transform and GRU neural network, which organically combines wavelet packet decomposition (WPT), variational mode decomposition (VMD) and GRU neural network to form an efficient and accurate island detection method, and can accurately identify the island working condition to provide protection for stable operation of the power grid.

[0007] The technical solution for achieving the above-mentioned object is: a photovoltaic power station island detection method based on wavelet transform and GRU neural network, comprising the following steps:

[0008] S1, using wavelet packet decomposition technology to filter the collected original voltage signal waveform of the photovoltaic power station, and filtering and removing the components with a frequency higher than 2000Hz in the original voltage signal;

[0009] S2, inputting the voltage signal filtered by the wavelet packet decomposition technology into a variational mode decomposition model, adaptively decomposing the filtered voltage signal into multiple modal components with different center frequencies through a variational mode decomposition algorithm, realizing preliminary extraction and separation of the internal characteristics of the voltage signal, and using a technology combining variational mode decomposition and information entropy to construct a voltage signal feature vector meeting the input requirements of the GRU neural network;

[0010] S3, inputting the decomposed modal component data into a GRU neural network model to complete the learning and training of the voltage signal feature data and the classification judgment, and finally realizing accurate detection and identification of the photovoltaic power station island phenomenon.

[0011] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S1, the wavelet packet decomposition takes into account both the time domain resolution and the frequency domain resolution of the voltage signal, and adaptively decomposes the input voltage signal at multiple levels according to actual requirements; after completing the signal decomposition, the components in the specific frequency range of the voltage signal are accurately filtered out by reconstructing the decomposed coefficients, and then the targeted filtering of the voltage signal is realized, laying a foundation for subsequent signal feature extraction.

[0012] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S1, the wavelet packet decomposition of the continuous time domain signal x(t) is mathematically represented as:

[0013] (1)

[0014] In formula (1), is a wavelet mother function with two variables of scale parameter a and translation parameter b; is The plural form of a; a is a scale factor, which performs scaling transformation on the wavelet; b is a translation factor, which performs translation transformation on the wavelet;

[0015] The wavelet packet decomposition reconstructs the high-frequency signal and the low-frequency signal decomposed by the inverse transformation, and the reconstruction formula is:

[0016] (2)

[0017] In formula (2), is a wavelet constant, which ensures that the wavelet transformation exists inverse transformation, so that the original signal is perfectly reconstructed; is a wavelet mother function; X(a, b) is a wavelet coefficient, which represents the component intensity of the signal at the scale parameter a and the translation parameter b; db is the differential of the translation parameter b; da is the differential of the scale parameter a;

[0018] Through the above steps, the decomposition and reconstruction of the time domain signal x(t) are completed, and when the reconstructed signal is filtered, the harmonics higher than 2000Hz are filtered according to the requirements of signal processing.

[0019] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S1, the wavelet packet decomposition selects coif wavelet, and carries out filtering processing on all collected original voltage signals.

[0020] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S2, the core target of the variational mode decomposition model is to find a set of mode sets with different modal center frequencies, and the corresponding variational problem is specifically described as follows: the original signal f is decomposed into k independent modal functions, under the constraint condition that the sum of each modal function is equal to the input signal f, the sum of the estimated bandwidth of each modal component is minimized through optimization calculation, and finally the efficient decomposition and feature extraction of the original signal are realized.

[0021] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S2, the energy entropy difference of the voltage at the common connection point before and after the island occurs is used to calculate the information entropy of each modal component decomposed by the variational mode decomposition model; through the change rule of each modal information entropy, the energy distribution change of the voltage at the common connection point in different frequency ranges before and after the island occurs is effectively reflected, which provides more distinguishable feature basis for accurate identification of island working condition.

[0022] The above-mentioned photovoltaic power station island detection method based on wavelet transform and GRU neural network, in step S2, suppose that the voltage signal filtered by the wavelet packet decomposition technology is decomposed into K modal function components, then the corresponding constraint variational mode decomposition model expression is:

[0023] (3)

[0024] (4)

[0025] In formula (3) and formula (4), is a Dirac distribution function; K is the number of decomposed modalities, }={ }is a set of each modal component, }={ is the center frequency corresponding to each modality, that is is the kth VMD decomposed modal component, is the center frequency of the kth modal component, is the voltage signal filtered by the wavelet packet decomposition technology, is the sum of each modal component of the variational modal decomposition model.

[0026] The photovoltaic power station island detection method based on wavelet transform and GRU neural network, the number of modalities K decomposed by the variational modal decomposition model is 4, the voltage signal filtered by the wavelet packet decomposition technology, 4 sub-signals are obtained by decomposing the voltage signal by the variational modal decomposition model, and the calculation method of the information entropy is:

[0027] (5)

[0028] In formula (5), is the ith modal component obtained by the variational modal decomposition, is the information entropy of the modality;

[0029] The four information entropies are taken as inputs of the GRU neural network model, and the four information entropies respectively represent modal energy in different frequency ranges; the obtained information entropy is normalized to reduce the dimension of the voltage signal, and the total information entropy calculation formula is:

[0030] (6)

[0031] In formula (6), E is the sum of the total four modal information entropies, E i is a single modal information entropy, and k is 4;

[0032] The single modal energy distribution proportion calculation formula is:

[0033] (7)

[0034] In formula (7), is the proportion of the ith modal component, E i is a single modal information entropy, and k is 4;

[0035] The normalized information entropy of each mode is combined to obtain a 1-row 4-column feature vector, and the feature vector is taken as the input of the GRU neural network model to learn the voltage features in different working conditions.

[0036] The photovoltaic power station island detection method based on wavelet transform and GRU neural network has the advantages that in step S3, the GRU neural network controls information flow through update gate and reset gate, has no independent memory unit, and only relies on hidden state to transmit information; the update gate determines the retention proportion of historical information; if the update gate value is close to 1, a large amount of historical state is retained; if the update gate value is close to 0, the current input is focused; the reset gate controls the influence of historical information on a new candidate state; if the reset gate is close to 0, the historical information is discarded, and the model is forced to focus on the new input; and the candidate hidden state is generated by combining the historical information filtered through the reset gate and the current input to generate a temporary state.

[0037] The photovoltaic power station island detection method based on wavelet transform and GRU neural network has the advantages that in step S3, the GRU neural network organizes different voltage feature vectors in the form of time series, and marks the island vector as 1 and the non-island vector as 0; if the GRU neural network output is 1, it represents the island working condition, and if the output is 0, it represents the non-island working condition.

[0038] The photovoltaic power station island detection method based on wavelet transform and GRU neural network fully considers the potential influence of the island detection process on the stable operation of the power grid, in order to avoid the interference of the detection operation on the normal working condition of the power grid, finally determines to adopt the island passive detection technology, organically combines wavelet packet decomposition (WPT), variational mode decomposition (VMD) and GRU neural network, forms an efficient and accurate island detection method, and can realize accurate identification of the island working condition. When the detection method of the application detects island abnormalities, the corresponding abnormal signal can be sent in time through the upper computer, and the stable operation of the power grid is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the photovoltaic power station island detection method based on wavelet transform and GRU neural network of the application;

[0040] Figure 2 The schematic diagram for decomposing and reconstructing voltage waveform data by using WPT;

[0041] Figure 3 The four mode condition diagrams obtained after the voltage signal is decomposed by VMD;

[0042] Figure 4 The structure and classification flowchart of the GRU neural network. DETAILED DESCRIPTION

[0043] In order for those skilled in the art of the technical field to better understand the technical solutions of the present application, the specific embodiments thereof will be described in detail below with reference to the accompanying drawings.

[0044] In the actual operation environment of the photovoltaic power station, the voltage and current signals under different working conditions (such as island, normal grid connection, load fluctuation, etc.) contain rich harmonic components, but the spectral characteristics of the two types of signals are significantly different, which is specifically manifested as follows: in terms of current signal, the current spectrum corresponding to each working condition lacks obvious distinguishing marks, and the energy distribution of different frequency harmonics presents a disordered state, if only relying on the current harmonic characteristics to construct a detection model, not only cannot effectively distinguish different working conditions, but also leads to low accuracy of island detection results, which is difficult to meet the engineering application requirements; and in terms of voltage signal, under various working conditions, the voltage spectrum of island working condition and non-island working condition has significant difference, and in the voltage spectrum corresponding to different working conditions, the energy distribution of each frequency component has clear regularity, which can provide clear feature support for working condition identification. Based on the above differences, the present application determines to take the voltage harmonic characteristics as the core judgment basis for island detection, for realizing accurate identification of island working condition.

[0045] Firstly, the wavelet packet decomposition (WPT) technology is adopted to process the collected original voltage signal waveform, and the components with frequency higher than 2000Hz in the signal are focused on and filtered out, so as to eliminate the influence of high-frequency interference signals on the subsequent detection link; secondly, the voltage signal filtered by WPT is input into the VMD module, and the filtered voltage signal is adaptively decomposed into multiple modal components with different center frequencies by the VMD algorithm, so as to realize preliminary extraction and separation of the internal characteristics of the voltage signal; finally, the related data of each modal component obtained by decomposition are input into the GRU neural network model, the advantages of GRU neural network in time series data learning and classification are utilized, the learning and training of voltage signal feature data and the classification judgment are completed, and finally the accurate detection and identification of island phenomenon are realized.

[0046] Referring to Figures 1 to 4 The embodiment of the present application is a photovoltaic power station island detection method based on wavelet transform and GRU neural network, which comprises the following steps:

[0047] S1, the wavelet packet decomposition (WPT) technology is adopted to filter and process the original voltage signal waveform of the photovoltaic power station collected, and the components with frequency higher than 2000Hz in the original voltage signal are focused on and filtered out, so as to eliminate the influence of high-frequency interference signals on the subsequent detection link;

[0048] S2, input the voltage signal filtered through the WPT into a variational mode decomposition (VMD) model, adaptively decompose the filtered voltage signal into a plurality of modal components with different center frequencies through the VMD algorithm, realize the preliminary extraction and separation of the internal characteristics of the voltage signal, and adopt the technology combining variational mode decomposition and information entropy to construct a voltage signal feature vector meeting the input requirements of the GRU neural network;

[0049] S3, input the related data of each modal component obtained by decomposition into a GRU neural network model, utilize the advantages of the GRU neural network in time series data learning and classification, complete the learning training and classification judgment of the voltage signal feature data, and finally realize the accurate detection and identification of the island phenomenon.

[0050] The application realizes deep feature mining of the voltage signal waveform through the multi-level cooperative processing mechanism, effectively filters out the interference components, and fully extracts the working condition difference characteristics, thereby guaranteeing the accuracy of island detection from the technical path, and meeting the requirement of detection accuracy for safe and stable operation of the photovoltaic power station.

[0051] (I) voltage signal filtering based on WPT:

[0052] The WPT model can simultaneously consider the time domain resolution and the frequency domain resolution of the voltage signal, can adaptively decompose the input voltage signal in multiple levels according to the actual requirement, and can accurately filter out the components in a specific frequency range of the voltage signal through reconstruction processing of the decomposed coefficients after completing signal decomposition, thereby realizing targeted filtering of the voltage signal and laying a foundation for subsequent signal feature extraction.

[0053] The wavelet packet decomposition of the continuous time domain signal x(t) is mathematically expressed as:

[0054] (1)

[0055] In formula (1), is a mother wavelet function, having two variables of a scale parameter a and a translation parameter b; is a complex form of ; a is a scale factor for scaling transformation of the wavelet; b is a translation factor for translation transformation of the wavelet.

[0056] The WPT can reconstruct the decomposed high-frequency and low-frequency signals through inverse transformation, and the reconstruction formula can be expressed as:

[0057] (2)

[0058] In formula (2), is a wavelet constant, which ensures that the wavelet transformation exists inverse transformation, so that the original signal is perfectly reconstructed; Ψ(a,b) is a mother wavelet function; X(a,b) is a wavelet coefficient, representing the component intensity of the signal at scale parameter a and translation parameter b; db is the differential of translation parameter b; da is the differential of scale parameter a.

[0059] Through the above steps, the decomposition and reconstruction of the time domain signal x(t) are completed. When reconstructing the signal, according to the requirements of signal processing, the harmonics higher than 2000Hz are filtered out.

[0060] WPT decomposition and reconstruction are equivalent to the combination of high-pass filter and low-pass filter. Using WPT to perform 3-layer decomposition on the voltage waveform can obtain multiple frequency bands in different frequency ranges.

[0061] Please refer to Figure 2 , WPT 3-layer decomposition of the voltage waveform can obtain 8 frequency bands: frequency band 1: 0-625Hz, frequency band 2: 625-1250Hz, frequency band 3: 1250-1875Hz, frequency band 4: 1875-2500Hz, frequency band 5: 2500-3125Hz, frequency band 6: 3125-3750Hz, frequency band 7: 3750-4375Hz, and frequency band 8: 4375-5000Hz. Selecting different frequency band signals for reconstruction can filter out harmonics in a specific frequency range.

[0062] Reconstruction is performed on the frequency components in the frequency band 1 to the frequency band 3, and a voltage signal with a frequency in the range of 0-1850Hz can be obtained. The remaining frequency bands (mainly frequency band 4 to frequency band 8) cover components with a frequency above 2000Hz, which are interference signals and need to be removed through filtering.

[0063] In the WPT processing process, the selection of the mother wavelet function directly affects the effect of signal reconstruction. Currently, commonly used mother wavelet functions in the industry include db wavelet, sym wavelet and coif wavelet. When different mother wavelet functions are applied to WPT filtering, the final filtering effect is different. Through special research and verification, it is found that, except for coif wavelet, the use of db wavelet, sym wavelet and other types of wavelet for signal reconstruction cannot completely eliminate the harmonic components with a frequency above 2000Hz, while coif wavelet performs best in filtering high-frequency harmonics and retaining effective signals, and can meet the filtering accuracy requirements of the voltage signal of the present application. Therefore, the present application determines to use coif wavelet to perform filtering processing on all collected original voltage data.

[0064] (2) Construction of voltage feature vector based on VMD:

[0065] The dimension of the voltage signal itself does not change after the denoising preprocessing is completed by the WPT. If the filtered voltage signal is directly used as the input data of the GRU neural network, the feature learning of the neural network will be significantly time-consuming due to the large amount of input data, which exceeds the time requirement of "island detection needs to be completed within 2 seconds" in the national standard. To solve this problem, technical means are needed to reduce the amount of data input into the GRU neural network, so as to improve the training efficiency and feature learning efficiency of the neural network.

[0066] The present application proposes a technical solution combining variational modal decomposition (VMD) and information entropy, which constructs a voltage feature vector that meets the input requirements of the GRU neural network. This method not only effectively compresses the data dimension and reduces the input data amount, but also accurately extracts the key features of the voltage signal, thereby improving the overall efficiency of island detection and ensuring that the detection process is completed within the time standard specified by the state.

[0067] The core goal of VMD is to find a set of modalities with different modal center frequencies. The corresponding variational problem can be described as follows: the original signal f is decomposed into k independent modal functions (IMF), and under the constraint condition that the sum of each modal function is equal to the input signal f, the sum of the estimated bandwidth of each modal component is minimized through optimization calculation, and finally the original signal is efficiently decomposed and features are extracted.

[0068] Suppose the voltage signal filtered by the wavelet packet decomposition technique is decomposed into K modal function components, then the corresponding constrained variational modal decomposition model expression is:

[0069] (3)

[0070] (4)

[0071] In formula (3) and formula (4), is the Dirac distribution function; K is the number of decomposed modes, }={ }is the set of each modal component, }={ }is the center frequency corresponding to each modal, that is is the kth VMD decomposed modal component, is the center frequency of the kth modal component, is the voltage signal filtered by the wavelet packet decomposition technique, is the sum of each modal component of the variational modal decomposition model.

[0072] The number of modes of the VMD solution needs to be determined in advance, the voltage signal filtered through the WPT is decomposed into K IMF components through VMD, the length of the IMF component is consistent with the original signal, if the number of modes of the decomposition is too much, the modes will be repeated and excessive noise will be generated, and the detection effect will be poor, if the number of modes of the decomposition is too small, the mode under-decomposition mode component cannot reflect the voltage waveform characteristics, and the detection accuracy will also be affected.

[0073] In the VMD decomposition process, the value of the mode number K has a key influence on the decomposition effect and the subsequent island detection accuracy. When the mode number K is less than 4, the center frequency of the mode component obtained by decomposition only covers the fundamental frequency and part of the harmonic frequency, and the characteristic information of the lower order harmonic frequency cannot be effectively captured, resulting in the loss of key working condition characteristics in the voltage signal; when the mode number K is greater than 4, the center frequency of different mode components will be close to each other in the decomposition process, and such overlapping characteristics will interfere with the subsequent working condition recognition and increase the risk of misjudgment.

[0074] Please refer to Figure 3 In order to simultaneously consider the complete extraction of the high-frequency component and the low-frequency component characteristics in the voltage signal, avoid feature loss or interference problems, the number of modes of VMD decomposition is finally determined to be 4, so as to ensure that the mode components after decomposition can fully and clearly reflect the inherent characteristics of the voltage signal, and provide reliable data support for subsequent feature vector construction.

[0075] In order to further improve the accuracy of island detection, the information entropy theory is introduced to describe the voltage feature vector, and the sparse characteristics of the voltage signal are quantitatively evaluated through information entropy. The numerical value of information entropy can directly reflect the uniformity of the probability distribution of signal feature parameters, and when the probability distribution is in the most uncertain state, the corresponding information entropy value reaches the maximum.

[0076] The present application utilizes the difference in energy entropy of the voltage at the point of common coupling (PCC) before and after the island occurs, and calculates the information entropy of each mode component obtained by VMD decomposition. Through the change rule of the information entropy of each mode, the energy distribution change of the PCC point voltage in different frequency ranges before and after the island occurs can be effectively reflected, and more distinguishable feature basis is provided for accurate identification of the island working condition.

[0077] Since the optimal mode number obtained by the variational mode decomposition is 4, the original voltage signal decomposition can obtain 4 sub-signals, and the calculation method of information entropy is:

[0078] (5)

[0079] In formula (5), is the i-th mode component obtained by variational mode decomposition, is the information entropy of the mode.

[0080] Four information entropies are obtained and used as inputs to the GRU neural network model. These four information entropies represent the modal energies of different frequency ranges. The obtained information entropies are normalized to reduce the dimensionality of the voltage signal. The formula for calculating the total information entropy is:

[0081] (6)

[0082] In equation (6), E is the sum of the entropies of the four modal information. i Let k be the entropy of a single modality, and k be 4.

[0083] The formula for calculating the proportion of energy distribution in a single mode is:

[0084] (7)

[0085] In equation (7), E represents the proportion of the i-th modal component. i Let k be the entropy of a single modality, and k be 4.

[0086] The normalized modal information entropies are combined to obtain a 1-row, 4-column feature vector, which is then used as the input to the GRU neural network model to learn and classify voltage characteristics under different operating conditions.

[0087] (III) Islanding detection of photovoltaic power plants based on GRU:

[0088] Traditional islanding detection methods suffer from large blind spots and low accuracy in practical applications, failing to meet the precision requirements of current photovoltaic power plants. Machine learning algorithms, with their powerful feature extraction capabilities, can capture subtle differences in voltage signals, providing an effective technical path to improve islanding detection accuracy.

[0089] The GRU neural network possesses strong learning performance, especially compared to the traditional BP neural network, exhibiting a more significant advantage in the field of time-series data recognition. The GRU neural network employs a time-series-based gradient descent method for learning and training. Its core mechanism, through built-in gating mechanisms (including reset and update gates), accurately captures the temporal dependencies of voltage feature vectors. During training, this mechanism dynamically adjusts the focus on historical and current features, progressively optimizing model parameters to minimize the target loss function, ultimately efficiently completing the classification task of voltage feature vectors under different operating conditions (islanded and non-islanded).

[0090] Please see Figure 4 The GRU neural network uses the update gate Z t With reset gatet Control information flow, no independent memory unit, only rely on hidden state transmission information; update gate Z t : Determine the retention ratio of historical information; if the update gate value is close to 1, a large number of historical states are retained; if the update gate value is close to 0, the current input is focused on. Reset gate r t : Control the influence of historical information on new candidate states; if the reset gate is close to 0, the historical information is discarded, and the model is forced to focus on new input. Candidate hidden state: generate temporary state by combining filtered historical information and current input with reset gate.

[0091] The present application organizes different voltage feature vectors in time series form, and marks the island vector as 1 and the non-island vector as 0. If the GRU neural network output is 1, it represents an island working condition, and if the output is 0, it represents a non-island working condition, so the output layer of the GRU neural network only needs 2 neurons, and the number of hidden layer neurons is set to 12.

[0092] In order to ensure the reliability of the training results of the GRU neural network, the ratio of training data to test data is 8:2. The back propagation process of the neural network uses gradient descent method to update the weight and threshold, and gradually finds the minimum value of the objective function to get the final model parameters. The selection of learning rate will affect the training speed of the neural network, and the learning rate is set to 0.0005.

[0093] Finally, through wavelet packet decomposition, VMD signal decomposition and GRU neural network, the island detection of photovoltaic power station is realized, and when the island anomaly is detected, the corresponding abnormal signal can be sent in time through the upper computer, which provides guarantee for the stable operation of power grid.

[0094] In summary, the photovoltaic power station island detection method based on wavelet transform and GRU neural network of the present application takes voltage harmonic features as the core judgment basis for island detection, organically combines wavelet packet decomposition (WPT), variational mode decomposition (VMD) and GRU neural network, forms a set of efficient and accurate island detection method, and can realize accurate identification of island working condition, which provides guarantee for stable operation of power grid.

[0095] Those skilled in the art in this technical field should realize that the above embodiments are only used to illustrate the present application, and are not used as a limitation on the present application, as long as the changes and modifications of the above described embodiments are within the scope of the spirit of the present application.

Claims

1. A method for detecting islanding in photovoltaic power plants based on wavelet transform and GRU neural network, characterized in that, Includes the following steps: S1, wavelet packet decomposition technology is used to filter the original voltage signal waveform of the photovoltaic power station, and the components with a frequency higher than 2000Hz in the original voltage signal are filtered out. S2, the voltage signal filtered by wavelet packet decomposition is input into the variational mode decomposition model. The filtered voltage signal is adaptively decomposed into multiple mode components with different center frequencies by the variational mode decomposition algorithm. The voltage signal feature vector that meets the input requirements of the GRU neural network is constructed by combining variational mode decomposition with information entropy. S3 inputs the decomposed modal component data into the GRU neural network model to complete the learning, training and classification of voltage signal feature data, and finally realizes the accurate detection and identification of islanding phenomenon in photovoltaic power stations.

2. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, In step S1, wavelet packet decomposition takes into account both the time-domain and frequency-domain resolution of the voltage signal. Based on actual needs, the input voltage signal is subjected to multi-level adaptive decomposition. After the signal decomposition is completed, the coefficients obtained by decomposition are reconstructed to accurately filter out components in a specific frequency range in the voltage signal, thereby achieving targeted filtering of the voltage signal and laying the foundation for subsequent signal feature extraction.

3. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, The wavelet packet decomposition of a continuous time-domain signal x(t) is mathematically represented as: （1） In equation (1), It is a wavelet mother function with two variables: scale parameter a and translation parameter b; yes The complex form of the wavelet; a is the scaling factor, which performs scaling transformation on the wavelet; b is the translation factor, which performs translation transformation on the wavelet; Wavelet packet decomposition reconstructs the high-frequency and low-frequency signals from the voltage signal using inverse transform. The reconstruction formula is as follows: （2） In equation (2), The wavelet constant ensures that the wavelet transform has an inverse transform, thus perfectly reconstructing the original signal; X is the wavelet mother function; X(a, b) are wavelet coefficients, representing the component intensity of the signal on the scale parameter a and the translation parameter b; db is the derivative of the translation parameter b; da is the derivative of the scale parameter a. The above steps complete the decomposition and reconstruction of the time-domain signal x(t). During signal reconstruction, harmonics with frequencies higher than 2000Hz are filtered out according to the requirements of signal processing.

4. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, In step S1, the wavelet packet decomposition uses the coif wavelet to perform filtering processing on all the acquired raw voltage signals.

5. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, In step S2, the core objective of the variational mode decomposition model is to find a set of modes with different modal center frequencies. The corresponding variational problem is specifically described as follows: decompose the original signal f into k independent mode functions. Under the constraint that the sum of the mode functions is equal to the input signal f, optimize the calculation to minimize the sum of the estimated bandwidths of each mode component, and finally achieve efficient decomposition and feature extraction of the original signal.

6. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 5, characterized in that, In step S2, the information entropy of each modal component obtained from the variational mode decomposition model is calculated by utilizing the difference in energy entropy of the voltage at the common connection point before and after islanding. By observing the change pattern of the information entropy of each mode, the energy distribution of the voltage at the common connection point before and after islanding is effectively reflected in different frequency ranges, providing more discriminative feature basis for the accurate identification of islanding conditions.

7. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 6, characterized in that, Suppose that the voltage signal filtered by wavelet packet decomposition is decomposed into K modal function components. Then the corresponding constrained variational mode decomposition model expression is: （3） （4） In equations (3) and (4), Let K be the Dirac distribution function; K is the number of modes in the decomposition, { }={ } represents the set of modal components, { }={ } represents the center frequency corresponding to each mode, i.e. For the k-th VMD decomposition mode component, Let be the center frequency of the k-th modal component. This is the voltage signal after being filtered using wavelet packet decomposition. It is the sum of the modal components of the variational mode decomposition model.

8. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 7, characterized in that, The variational mode decomposition model decomposes the voltage signal into four modes, K, after filtering with wavelet packet decomposition. The resulting signal is then further decomposed into four sub-signals. The information entropy is calculated as follows: （5） In equation (5), Let i be the i-th mode component obtained from variational mode decomposition. This represents the information entropy of the mode. Four information entropies are obtained and used as inputs to the GRU neural network model. These four information entropies represent the modal energies of different frequency ranges. The obtained information entropies are normalized to reduce the dimensionality of the voltage signal. The formula for calculating the total information entropy is: （6） In equation (6), E is the sum of the entropies of the four modal information. i The entropy is for a single modal information, and k is 4; The formula for calculating the proportion of energy distribution in a single mode is: （7） In equation (7), E represents the proportion of the i-th modal component. i The entropy is for a single modal information, and k is 4; The normalized modal information entropies are combined to obtain a 1-row, 4-column feature vector, which is then used as the input to the GRU neural network model to learn and classify voltage characteristics under different operating conditions.

9. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, In step S3, the GRU neural network controls the flow of information through update gates and reset gates. It has no independent memory unit and relies solely on hidden states to transmit information. Update gate: determines the proportion of historical information retained. If the update gate value is close to 1, a large amount of historical states are retained. If the update gate value is close to 0, the focus is on the current input. Reset gate: controls the influence of historical information on new candidate states. If the reset gate value is close to 0, historical information is discarded, forcing the model to focus on new inputs. Candidate hidden state: combines the historical information filtered by the reset gate with the current input to generate a temporary state.

10. The method for detecting islanding in a photovoltaic power station based on wavelet transform and GRU neural network according to claim 1, characterized in that, In step S3, the GRU neural network organizes different voltage feature vectors in the form of time series and marks islanded vectors as 1 and non-islanded vectors as 0; if the output of the GRU neural network is 1, it represents an islanded operating condition, and if the output is 0, it represents a non-islanded operating condition.

Images