A method, system, terminal and medium for detecting islanding of grid-connected inverter

Abstract

The invention discloses a method, system, terminal and medium for detecting islands of grid-connected inverters, relates to the field of microgrid detection technology, and solves the problems of large detection blind area and low detection accuracy in the prior art when detecting islands. The key points of the technical scheme are: using sliding window discrete Fourier transform and symmetrical component method to extract zero, positive and negative sequence feature quantities of harmonics in common coupling point (PCC) voltage and inverter current signal, using the extracted zero, positive and negative sequence components of harmonics as main feature quantities, system frequency as auxiliary feature quantity, and a total of 7 feature quantities to form the feature vector space of LSTM neural network for island detection. The purpose of accurate detection is achieved when the active power and reactive power of the power supply and local load in the microgrid are basically matched, and the purpose of accurate identification of islands and high-precision detection is achieved when the PCC point circuit breaker trips, three-phase short circuit and local load mutation occurs. In addition, it has the advantages of optimizing the island feature extraction method and not affecting the power quality.

Description

Technical Field

[0001] The present invention relates to the field of microgrid detection technology, and more specifically, to a method, system, terminal and medium for detecting an island of a grid-connected inverter. Background Art

[0002] With the rapid development of the photovoltaic power generation industry, the grid-connected photovoltaic inverters will have a certain impact on the stability of the power grid. The islanding phenomenon refers to the situation when the power grid is abnormal, and the various distributed power systems fail to detect the power outage in time, and continue to provide power to the local load, forming a power supply system composed of the power generation system and the load. It is an unplanned island and is defined as a fault. When an unplanned island occurs, the reliability of the power supply cannot be guaranteed, which damages electrical equipment and even endangers the personal safety of operation and maintenance personnel. In order to ensure safety and power quality, according to GB / T33593-2017 and IEEE 1547, when operating in an unplanned island, the microgrid must have the ability to quickly detect islands, and the maximum detection time shall not exceed 2s. Therefore, effective islanding detection is a basic requirement for microgrids.

[0003] At present, the methods of island detection can be divided into three categories: remote method, active method and passive method. The remote detection method realizes detection through signal acquisition and communication technology. The disadvantages of this method are high detection cost, complexity and difficulty in implementation; the active detection method introduces disturbance signals to cause the amplitude or frequency of the output voltage / current of the distributed power supply to deviate from the normal range, thereby triggering the relay to work to determine the island state, such as frequency shift method, power disturbance method, impedance measurement method, etc. This method will not cause obvious distortion of the output voltage of the distributed power supply under the grid-connected operation of the inverter, but will cause power quality problems and harmonic interference, thereby reducing system performance; the passive detection method determines the island based on the measurement of abnormal voltage or frequency, such as over / under voltage detection method, over / under frequency detection method, voltage harmonic detection method, etc. This method has a large detection blind area when the power is matched, and it is difficult to select the detection threshold. The detection method combining active and passive methods is both complex and inefficient. When the active power and reactive power do not match, the existing blind area and threshold setting are the main factors affecting the detection.

[0004] In view of the shortcomings of traditional island detection methods, some people have proposed detection methods based on signal processing and artificial intelligence to overcome the shortcomings of existing methods. Among them, the island detection method based on BP neural network only uses the voltage value within the cycle as the input value of the BP neural network, which is prone to local optimal solutions. The intrinsic mode function detects non-stationary modal voltage signals, which is prone to false detection when the power is matched. The support vector machine extracts the voltage waveform at the distributed power supply end to identify the island. This method does not consider multiple types of distributed power supplies. A distributed power supply island detection method based on LSTM uses the LSTM neural network as a feature extractor and classifier, resulting in a great influence of the LSTM model structure on the detection accuracy. Too many or too few model layers will cause false detection. Wavelet transform extracts the complex time frequency characteristics of the PCC point voltage signal, and uses the LSTM neural network for classification to achieve accurate detection of the island state. These methods can accurately classify island and non-island events and detect islands when there is low power imbalance. However, none of them considers that when the active power and reactive power output by the inverter do not change before and after the islanding occurs, the voltage and frequency of the PCC point in the islanding mode will change very little, and it is difficult to detect the islanding state in this case. Summary of the invention

[0005] The purpose of the present invention is to provide a grid-connected inverter island detection method, system, terminal and medium to solve the problem that when the active power and reactive power output by the inverter do not change before and after the islanding occurs, the voltage and frequency of the PCC point in the islanding mode change very little, making it difficult to detect the islanding state.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions:

[0007] A method for detecting an islanding of a grid-connected inverter comprises the following steps:

[0008] Obtain PCC point voltage, inverter output current and system frequency;

[0009] The zero, positive and negative sequence characteristics of harmonics in the PCC point voltage and inverter output current are extracted by using sliding window discrete Fourier transform and symmetric component method, and six main eigenvectors are obtained.

[0010] The system frequency is used as the auxiliary eigenvector and combined with the six main eigenvectors to divide the training set and the test set;

[0011] An LSTM classifier model is constructed, and the training set is used as the feature vector space of the LSTM classifier model to perform island detection and obtain the detection results.

[0012] Furthermore, the acquisition process of the PCC point voltage, the inverter output current and the system frequency is specifically as follows:

[0013] Build a photovoltaic inverter simulation model through MATLAB / Simulink;

[0014] Set the grid parameters and simulate the operation of the system under different load quality factors, such as main circuit breaker tripping, PCC three-phase short circuit and local load mutation, under island and non-island conditions;

[0015] The sampling frequency is set to 5kHz to collect the PCC point voltage, inverter output current and system frequency.

[0016] Furthermore, the extraction process of the main feature vector is specifically as follows:

[0017] The SDFT algorithm is used to detect the harmonics of the PCC point voltage and the inverter output current;

[0018] Determine the harmonic that has the greatest impact on the system among all harmonics;

[0019] The symmetrical component method is used to separate the most influential harmonics, and six principal eigenvectors of the PCC point voltage and inverter output current corresponding to the zero-sequence component and the positive-sequence and negative-sequence components are obtained.

[0020] Further, the LSTM classifier model is configured with an input layer, a first LSTM layer and a second LSTM layer;

[0021] Input layer, used to input time series into the network;

[0022] The first LSTM layer is used to generate the corresponding hidden state and updated cell state through the initial state and input sequence;

[0023] The second LSTM layer is used to update the current cell state and calculate the output hidden state and another updated cell state.

[0024] Furthermore, the hidden state in the LSTM classifier model is connected to the softmax layer through a fully connected layer;

[0025] The LSTM classifier uses a softmax layer for classification;

[0026] The softmax layer maps the results to probability values ​​between 0 and 1 and makes the sum of all output probability values ​​equal to 1.

[0027] Furthermore, the probability calculation formula of the softmax layer is specifically:

[0028] Y = softmax(W s h f +b s )

[0029]

[0030] Among them, Y is the probability, W S , b s is the learning parameter, h f is the hidden state, C represents the fault type, Z i represents the value of the i-th fault, Z j Represents the value of the jth fault.

[0031] Furthermore, the specific process of the LSTM classifier model for island detection is as follows:

[0032] The system frequency and the six main eigenvectors are standardized, and the standardized system frequency and the six main eigenvectors are divided into a training set and a test set;

[0033] Initialize the LSTM model and use the training set as the feature vector space of the LSTM neural network. After the training is completed, export the model parameters of the offline training as the parameters of the online detection model.

[0034] The recognition accuracy of the online detection model is tested through the test set, the model performance is evaluated, and the performance of the proposed algorithm is verified under different working conditions.

[0035] A grid-connected inverter island detection system, comprising:

[0036] Data acquisition module, used to obtain PCC point voltage, inverter output current and system frequency;

[0037] A feature extraction module is used to extract zero-sequence, positive-sequence and negative-sequence feature quantities of harmonics in the PCC point voltage and the inverter output current by using a sliding window discrete Fourier transform and a symmetric component method to obtain six main feature vectors;

[0038] The feature division module is used to divide the training set and the test set after taking the system frequency as the auxiliary feature vector and combining it with the six main feature vectors;

[0039] The model training module is used to build an LSTM classifier model, and use the training set as the feature vector space of the LSTM classifier model to perform island detection and obtain detection results.

[0040] An electronic terminal, comprising:

[0041] A memory, the memory being used to store a computer program;

[0042] A processor is used to execute the computer program stored in the memory so that the electronic terminal executes the above-mentioned grid-connected inverter island detection method.

[0043] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned grid-connected inverter island detection method.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] ① The sliding window discrete Fourier transform method is used to detect the harmonics of the common coupling point voltage and inverter output current before and after the islanding occurs, and determine the harmonic that has the greatest impact on the system. The purpose of accurate detection is achieved when the active power and reactive power are basically matched, avoiding the problems of large blind areas and difficult threshold setting in passive detection.

[0046] ② The zero, positive and negative sequence components of the harmonics that have the greatest impact on the system are extracted using the symmetrical component method, and the island feature extraction method is optimized.

[0047] ③ The zero, positive and negative sequence components of the harmonics that have the greatest impact on the system and the system frequency are used as the main feature quantities, and the LSTM neural network model is used to obtain the detection results. This method introduces disturbance signals and will not cause power quality problems.

[0048] ④ Use MATLAB / Simulink to build a photovoltaic inverter simulation model, set the grid parameters, and simulate the system under different quality factors, PCC point circuit breaker tripping, three-phase short circuit and local load mutation. This method can still accurately identify islands with high detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0050] Figure 1 A schematic diagram of the structure of a grid-connected inverter system according to an embodiment of the present invention;

[0051] Figure 2 It is a flow chart of the present invention using the symmetrical component method to obtain the characteristic quantity of the harmonic that has the greatest impact on the system;

[0052] Figure 3 This is a schematic diagram of the LSTM framework in an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of the specific mechanism structure of LSTM cell in an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of a novel island detection model provided by the present invention;

[0055] Figure 6The accuracy curve and loss curve of the trained model, where (a) is the accuracy curve and (b) is the loss curve;

[0056] Figure 7 The SDFT extraction result in Example A in the embodiment;

[0057] Figure 8 : It is the island simulation waveform with a quality factor of 1 in Example A in the embodiment, where (a) is the grid current, (b) is the system frequency, (c) is the PCC voltage, (d) is the positive and negative zero-sequence components of the PCC voltage, (e) is the inverter output current, (f) is the positive and negative zero-sequence components of the inverter output current, and (g) is the island identification result. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0059] Embodiment: A method, system, terminal and medium for detecting islanding of a grid-connected inverter.

[0060] Among them, the sliding window discrete Fourier transform (SDFT) is evolved from the discrete Fourier transform (DFT), and the calculation formula of the discrete Fourier transform is:

[0061]

[0062] Where N is the calculation length, x(n) is the nth sampling value, k is the harmonic number, and X(k) is the detection result of the kth harmonic. The sliding window discrete Fourier transform adds a sliding sampling window on the basis of the discrete Fourier transform. In every two adjacent calculations, the program only updates one sampling value. The calculation formula of the sliding window discrete Fourier transform is:

[0063]

[0064] Where x(n) is the current sampling value, X S (k-1) is the result of the last calculation.

[0065] like Figure 1 The structure of the grid-connected inverter system is shown in Figure 1. A transformer and a circuit breaker CB1 are connected to the grid, where the power output of the inverter is P L +jQ L , the power generated by the grid is PL +jQ L , the power absorbed by the load is P L +jQ L .

[0066] The voltage of the point of common coupling (PCC) and the output current of the inverter are collected. SDFT detects the discrete spectrum of the harmonics in the PCC voltage and the inverter output current. The amplitude and frequency information of each harmonic can be accurately obtained, the interference in the PCC voltage and the inverter current can be determined, and the islanding event can be further determined.

[0067] Since the harmonics detected by SDFT are very complex, including positive sequence components, negative sequence components, and zero sequence components, and the zero sequence components have no effect on the system in the experiment, it is necessary to use the Symmetrical Components Method (SCM) to extract the positive and negative zero sequence components and separate the zero sequence components from the feature matrix. The size of the symmetrical sequence component is the main indicator for indicating the presence of interference in the voltage or current waveform. The three-phase voltage (V a 、V b 、V c ) and the inverter output current (I a ,I b ,I c ) is decomposed symmetrically to obtain the following equation:

[0068]

[0069]

[0070] Where a is an operator, and a=1∠120 0 ; V p 、V n , V0 are the positive, negative and zero sequences of voltage respectively; I p ,I n , I0 are the positive, negative and zero sequences of current respectively.

[0071] Figure 2 This is the feature extraction step of the proposed island detection. The SDFT algorithm detects the harmonics of the PCC point voltage and the inverter output current, determines the harmonic that has the greatest impact on the system, and then the symmetrical component method separates the zero-sequence component from the positive-sequence and negative-sequence components.

[0072] The LSTM neural network model consists of an input layer and an LSTM layer. The input layer inputs time series data into the network, and the LSTM layer learns the correlation between sequence data and time. The LSTM framework for island detection is shown in Figure 1. Figure 3As shown in the figure, a series of time series pass through the LSTM neural network. The first LSTM cell passes through the initial state (C t-1 ,h t-1 ) and the input sequence (x 11 , x 21 ..x k1 ) generates the corresponding hidden state (h 1t ,h 2t ...h kt ) and the updated cell state (C t ), the next LSTM cell updates the current cell state and calculates the output hidden state and another updated cell state. Where x is the input signal with k features, and N is the number of hidden layer units for island classification.

[0073] The specific mechanism of LSTM cell is as follows Figure 4 As shown in the figure, the network structure is designed with three main gate units with gate activation function. In the figure, σ is the sigmoid activation function, which can map real numbers to [0, 1]. 0 means that all the information of the unit at the previous moment is discarded, and 1 means that all the information is retained; tanh is the hyperbolic tangent function, which can map real numbers to [-1, 1].

[0074] where f t For the forget gate, i t is the input gate, o t The output gate and the update gate unit are as follows

[0075]

[0076] Where U∈R d×d , W∈R d×k , b∈R d is a learning parameter, k, d are the input size and the size of the hidden layer respectively, h t-1 , X t are the output at the previous moment and the input at the current moment respectively. In island detection, the input signal X t Represents the PCC point voltage, the positive, negative, and zero-sequence components of the inverter current, and the system frequency. Island detection is essentially a binary classification. f contains most of the information in the input signal, so we will f The fully connected layer is connected to the softmax layer, and the softmax layer is used for classification. The softmax layer maps the result to a probability value between 0 and 1, and makes the sum of all output probability values ​​​​1. The probability is calculated as follows, where Y is the probability, W s ∈R |C|×d , b s ∈R |C|is the learning parameter, C represents the fault type, Z i represents the value of the i-th fault, Z j Represents the value of the jth fault.

[0077] Y = softmax(W s h f +b s ) (6)

[0078]

[0079] Island detection is a binary classification problem, that is, the output discrimination result consists of two states, 1 represents an island event and 0 represents a non-island event. The new island detection model based on SDFT-LSTM is as follows: Figure 5 As shown, it is mainly divided into offline training and online detection. The specific steps are as follows:

[0080] Step 1: Use MATLAB / Simulink to build a PV inverter simulation model, set grid parameters, and simulate the operation of the system under different load quality factors, such as main circuit breaker tripping, PCC three-phase short circuit, and local load mutation, under island and non-island conditions. Set the sampling frequency to 5kHz, and collect the PCC point voltage, inverter output current, and system frequency.

[0081] Step 2: Obtain the amplitude and frequency information of each harmonic through sliding window discrete Fourier transform, and determine the harmonic that has the greatest impact on the system.

[0082] Step 3: Use the symmetrical component method to calculate the positive, negative and zero sequence components of the harmonics, standardize the PCC point voltage, the positive, negative and zero sequence components of the inverter output current and the system frequency, and divide them into training set and test set.

[0083] Step 4: Initialize the LSTM model and use the training set as the feature vector space of the LSTM neural network. After the training, export the model parameters of the offline training as the parameters of the online detection model.

[0084] Step 5: Use the test set to test the recognition accuracy of the online detection model, evaluate the model performance, and verify the performance of the proposed algorithm under different working conditions.

[0085] Case analysis:

[0086] Test system

[0087] A photovoltaic inverter grid-connected simulation model was built in MATLAB / Simulink. A 690V / 30MVA three-phase voltage source was connected to a 690V / 270V step-down transformer, which was connected to the photovoltaic inverter through a three-phase circuit breaker. The system power frequency was 50Hz, and the inverter switching frequency was 5kHz. The specific simulation parameters are shown in Table 1. Before and after the islanding occurs, the deviation of active power and reactive power is approximated to 0. At this time, the system has the greatest difficulty in detecting the islanding.

[0088]

[0089] Table 1

[0090] LSTM offline training model training

[0091] The parameters and hyperparameters of the LSTM offline training model are optimized based on multiple experiments and training. The specific values ​​are shown in Table 2.

[0092]

[0093] Table 2

[0094] The accuracy curve and loss curve of the offline training model after training are as follows Figure 6 At this time, the weight parameters of the offline training model are exported and used for the online recognition model, and its accuracy on the test set can theoretically reach 100%.

[0095] Verification and analysis of different working conditions

[0096] The load quality factor is used to measure the resonance capability of the local load. The relationship between the two is proportional. The larger the load quality factor, the stronger the local load resonance capability. When the operating frequency of the system is equal to the resonance frequency, the active and reactive power output by the inverter basically does not change before and after the circuit breaker is disconnected, which will cause the voltage change of the grid-connected point in the islanding mode to be very small, and it is difficult to detect the islanding fault. In order to verify the performance of the proposed algorithm under different working conditions, the trained offline training model is used for islanding and non-islanding events: PCC point circuit breaker tripping with quality factors of 1 and 2.5 respectively; PCC point three-phase short circuit; local load mutation event.

[0097] Example A: PCC point circuit breaker tripping with quality factor 1

[0098] The parameters of the parallel RCL load are: R = 0.64Ω, C = 4.97mF, L = 2.04mH. At this time, the load quality factor is 1. The three-phase circuit breaker is set to disconnect at 0.8s. The amplitude of each harmonic is detected by SDFT. Figure 7 It can be seen that the second harmonic has a greater impact on islanding events. Figure 8(c) and (e) There is no obvious change in the inverter voltage and current before and after the circuit breaker is disconnected at 0.8s, and the system frequency is stable. At this time, it is difficult for the system to detect the island. Figure 8 This is the island simulation waveform when the circuit breaker CB1 is disconnected at 0.8s. Figure 8 (a)(b) shows that the system grid current is 0 after 0.8s. The positive and negative zero-sequence components of the PCC point voltage and inverter current show the interference in the voltage and current waveforms. Between 0 and 0.8s, there is no islanding event in the system, and its component is zero. After 0.8s, the component undergoes a significant mutation. Both the over / under voltage and over / under frequency detection methods cannot identify. Figure 8 The model shown in (g) identifies the tripping of the PCC point circuit breaker with a quality factor of 1 as an island.

[0099] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting an island in a grid-connected inverter, characterized in that: The following steps are involved: Obtain PCC point voltage, inverter output current and system frequency; Use sliding window discrete Fourier transform algorithm to detect the harmonics of PCC point voltage and inverter output current; determine the harmonic with the greatest impact on the system among all the harmonics; The symmetrical component method is used to separate the most influential harmonics, and six principal eigenvectors of the PCC point voltage and inverter output current corresponding to the zero-sequence component, positive-sequence component and negative-sequence component are obtained. The system frequency is used as the auxiliary eigenvector and combined with the six main eigenvectors to divide the training set and the test set; An LSTM classifier model is constructed, and the training set is used as the feature vector space of the LSTM classifier model to perform island detection and obtain the detection results.

2. A method for detecting an islanding of a grid-connected inverter according to claim 1, characterized in that: The acquisition process of the PCC point voltage, inverter output current and system frequency is specifically as follows: Build a photovoltaic inverter simulation model through MATLAB / Simulink; Set the grid parameters and simulate the operation of the system in island and non-island conditions with main circuit breaker tripping, PCC three-phase short circuit and local load mutation under different quality factors; The sampling frequency is set to 5kHz to collect the PCC point voltage, inverter output current and system frequency.

3. The method for detecting an islanding of a grid-connected inverter according to claim 1, characterized in that: The LSTM classifier model is configured with an input layer, a first LSTM layer, and a second LSTM layer; Input layer, used to input time series into the network; The first LSTM layer is used to generate the corresponding hidden state and updated cell state through the initial state and input sequence; The second LSTM layer is used to update the current cell state and calculate the output hidden state and another updated cell state.

4. The method for detecting an islanding of a grid-connected inverter according to claim 1, characterized in that: The hidden state in the LSTM classifier model is connected to the softmax layer through a fully connected layer; The LSTM classifier uses a softmax layer for classification; The softmax layer maps the results to probability values ​​between 0 and 1 and makes the sum of all output probability values ​​equal to 1.

5. A method for detecting an islanding of a grid-connected inverter according to claim 4, characterized in that: The probability calculation formula of the softmax layer is specifically: Y =softmax(W s h f +b s ) Among them, Y is the probability, W S , b s is the learning parameter, h f is the hidden state, C represents the fault type, Z i represents the value of the i-th fault, Z j Represents the value of the jth fault.

6. The method for detecting an islanding of a grid-connected inverter according to claim 1, characterized in that: The specific process of island detection by the LSTM classifier model is as follows: The system frequency and the six main eigenvectors are standardized, and the standardized system frequency and the six main eigenvectors are divided into a training set and a test set; Initialize the LSTM model and use the training set as the feature vector space of the LSTM neural network. After the training is completed, export the model parameters of the offline training as the parameters of the online detection model. The recognition accuracy of the online detection model is tested through the test set, the model performance is evaluated, and the performance of the proposed algorithm is verified under different working conditions.

7. A grid-connected inverter island detection system, characterized in that: include: Data acquisition module, used to obtain PCC point voltage, inverter output current and system frequency; A feature extraction module is used to detect the harmonics of the PCC point voltage and the inverter output current using a sliding window discrete Fourier transform algorithm; and to determine the harmonic that has the greatest impact on the system among the harmonics; The symmetrical component method is used to separate the most influential harmonics, and six principal eigenvectors of the PCC point voltage and inverter output current corresponding to the zero-sequence component, positive-sequence component and negative-sequence component are obtained. The feature division module is used to divide the training set and the test set after taking the system frequency as the auxiliary feature vector and combining it with the six main feature vectors; The model training module is used to build an LSTM classifier model, and use the training set as the feature vector space of the LSTM classifier model to perform island detection and obtain detection results.

8. An electronic terminal, characterized in that: include: A memory, the memory being used to store a computer program; A processor, wherein the processor is used to execute the computer program stored in the memory so that the electronic terminal executes a grid-connected inverter island detection method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, a grid-connected inverter island detection method according to any one of claims 1 to 6 is implemented.

Images