SVG-containing wind field high-frequency resonance online identification method based on fuzzy cognitive map

Abstract

The invention discloses an SVG-containing wind field high-frequency resonance online identification method based on a fuzzy cognitive map, and belongs to the technical field of power system identification. The invention aims to provide an SVG-containing wind field high-frequency resonance online identification method based on a fuzzy cognitive map for performing wind field high-frequency resonance identification on SVG outlet voltage and current time sequences acquired in a period. According to the method, SVG outlet voltage and current in a current inner loop of a controller of the SVG outlet voltage are selected in a wind field, amplitude-frequency characteristics of the current in the current inner loop are extracted through fast Fourier transform in secondary identification, and interval data of each time window and the amplitude-frequency characteristics of each time window are fuzzified through fuzzy mean clustering. And performing data labeling on the obtained membership matrix, and inputting the data into a resonance identification model for identification testing. According to the SVG electrical detection signal, high-frequency resonance identification can be rapidly completed in advance, and technical support is provided for timely starting SVG control mode switching.

Description

Technical Field

[0001] The invention belongs to the technical field of power system identification. Background Art

[0002] Existing wind farms are generally equipped with static VAR generators (SVG) to achieve dynamic reactive power regulation to meet the control requirements of regional system reactive voltage. However, there is an interaction between SVG and grid impedance, wind turbines and collector cables, which may cause high-frequency resonance of the system and affect the stable operation of the system. Both SVG and DFIG converters are fast-response power electronic converters. Under certain working conditions, the electrical (control) coupling between modules will increase the risk of high-frequency resonance of the system. For example: In 2021, photovoltaic power stations containing SVG in the Huyanghe area frequently experienced high-frequency resonance of about 2500Hz, causing serious overvoltage of the busbar and damage to a large number of grid-connected power electronic equipment; in May of the same year, the Wudalai wind farm in northern Hebei experienced about 50 resonances, resulting in serious distortion of the grid-connected busbar voltage waveform and SVG tripping due to excessive locking times.

[0003] Domestic and foreign scholars have conducted a lot of research on the resonance identification of wind farm systems. Traditional methods for extracting resonance data features mainly include time domain, frequency domain, time-frequency domain, etc., such as empirical mode decomposition (EMD) wavelet transform, power spectrum entropy, etc.; traditional identification models include BP neural network, probabilistic neural network (PNN), support vector machine (SVM), convolutional neural network (CNN) resonance identification model, etc. However, the research focus of these models is more on improving the identification accuracy of resonance. However, in actual engineering applications, the identification model is more required to have rapid judgment and strong robustness to prevent the expansion of system resonance problems. Therefore, it is necessary to form an online high-frequency resonance identification method to promptly detect high-frequency resonance problems at the beginning of high-frequency resonance in wind farm systems containing SVG, and provide technical support for subsequent SVG control mode switching (suppressing resonance phenomena). Summary of the invention

[0004] The purpose of the present invention is to provide an online identification method of high-frequency resonance of a wind farm containing SVG based on fuzzy cognitive graph for identifying high-frequency resonance of the wind farm according to the time series of SVG outlet voltage and current collected within one cycle.

[0005] The steps of the present invention are: S1, SVG outlet voltage U s And the current I in the controller current inner loop s_d ,I s_q The time series X={x1,x2,x3,......,x t} indicates that its length is t. The time series is divided into k windows with one cycle as the time window. The time length of each window is w. Then the relationship between k, t, and w is k = t / w. The time series after segmentation is: During the first identification, the SVG output voltage U s The time series is divided into windows, and each time window is fuzzy processed using a triangular fuzzy set to obtain a fuzzy information granule G representing each window, and finally a granular time series G = {G1, G2, ..., G K}, as shown in Formula 2: Among them, the granulation interval G K = {Low k ,R k ,Up k}, R is the average value of the change in the time window data, Low and Up represent the lower and upper limits of the interval coverage amplitude respectively; During the secondary identification, the current I in the inner loop of the controller is s_d ,I s_q Each time window of is decomposed by fast Fourier transform to obtain the amplitude-frequency characteristic matrix F = {F1, F2, ..., F K} T , as shown in Formula 3: Among them, the amplitude F of each time window k ={F k1 ,F k2 ,...,F kw}; The information particle G or the amplitude-frequency characteristic F of each time window is regarded as a data sample. After clustering the samples by fuzzy C-means clustering, the membership matrix U = {U1,U2,...,U K} T , the membership matrix is ​​shown in Formula 4: Among them U k = {U k1 ,U k2 ,...,U Kc}, c is the number of clusters of fuzzy C-means; Fuzzy cognitive graphs of different sizes are established according to the number of clusters. The membership matrix U is used as learning data, and each learning data is labeled A = {A1,...,A K} T , where A={A11 ,...,A 1i}, i is the number of sample classification categories, and the convex optimization algorithm is used to learn the fuzzy cognitive map, as shown in Formula 5: A i (t) = f(U(t)w i ) (5) Where U(t)=[U1,U2,...,U c ], t=1,2,......,k is the state value of all nodes of FCM at time t, A i (t) is the system state value of the i-th node of FCM at time t, W i = {W 1i ,W 2i ,...,W ci} T is the i-th column of the weight matrix; Formula 5 can be transformed by inverse transformation: f -1 (A i (t))=U(t)w i (6) According to formula 6, the following linear equation can be generated: Y i =UW i St|W ji ∣≤1,j=1,2,…,C (7) Objective function: min||U*w i -Y i ||2+β||w i ||1 st||w i || ∞ ≤1 (8) Repeat the above steps to learn the weight matrix of all time series, and the fuzzy cognitive map completes the learning of the weight matrix W; S2, the recognition accuracy of the identifier running with the fuzzy cognitive map system: Among them, X T is the identification sample base; X error Identify the number of incorrect samples; The identification error is:

[0006] The present invention does not need to detect the operating status of numerous wind turbines in the wind farm. Before the wind farm protection device responds, the high-frequency resonance can be quickly identified in advance based on the SVG electrical detection signal, thereby providing technical support for timely starting the SVG control mode switching. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A topological connection diagram of a wind farm grid-connected system provided in an embodiment of the present invention; Figure 2 A current inner loop structure diagram of an SVG control system provided in an embodiment of the present invention; Figure 3a A fuzzy cognitive map provided in an embodiment of the present invention; Figure 3b A schematic diagram of the reasoning process provided in an embodiment of the present invention; Figure 4 It is an overall architecture diagram of the fuzzy cognitive graph model provided in an embodiment of the present invention; Figure 5 A flowchart of the double resonance online identifier provided in an embodiment of the present invention; Figure 6 A flow chart of a resonance identification modeling process provided in an embodiment of the present invention; Figure 7a is the original outlet voltage U of SVG provided in the embodiment of the present invention s Waveform graph; Figure 7b The d-axis component I of the original outlet current of SVG provided in the embodiment of the present invention s_d Waveform graph; Figure 7c is the q-axis component I of the original outlet current of SVG provided in the embodiment of the present invention s_q Waveform graph; Figure 8a is the SVG outlet voltage U provided in the embodiment of the present invention s Data processing diagram; Figure 8b is the SVG outlet current I provided in the embodiment of the present invention s_d and I s_q Data processing diagram under strong working conditions; Figure 8c is the SVG outlet current I provided in the embodiment of the present invention s_d and I s_q Data processing diagram under weak working conditions; Figure 9a An offline primary resonance identification error diagram provided in an embodiment of the present invention; Figure 9b An offline secondary resonance identification error diagram provided in an embodiment of the present invention; Fig.10 The following are effect diagrams of two real-time resonance identifications provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0008] In order to overcome the shortcomings of low identification accuracy, slow identification speed and poor robustness of the current traditional resonance identification model, the present invention provides an online identification method for high-frequency resonance of a wind farm system containing SVG based on a fuzzy cognitive map, comprising: Using granular computing and FFT, Figure 1 The wind farm grid-connected system SVG output voltage U s As well as Figure 2 The d-axis component of the inner loop current in the SVG controller shown in FIG. s_d , q-axis component I s_q High-frequency resonance features are extracted, and the membership matrix is ​​obtained through fuzzy C-means clustering. This matrix is ​​used as a training sample to complete the training of the fuzzy cognitive graph recognition model and obtain the resonance recognition model. The model is embedded in the SVG control system to realize online identification of high-frequency resonance in the wind farm, providing technical support for SVG control mode switching (resonance suppression).

[0009] In the embodiment of the present invention, the resonance identification SVG outlet voltage and current time series processing algorithm based on fuzzy cognitive graph includes: When processing voltage time series data sets, one-time identification is performed to convert the numerical voltage time series into interval voltage time series through granular computing. Granular computing takes each cycle as a time window, applies reasonable granularity principles to each time window, and obtains appropriate numerical intervals. As a preliminary identification, one-time identification will significantly reduce the amount of model calculation and greatly improve the identification speed.

[0010] In order to improve the accuracy of identification and prevent misjudgment in the first identification, the present invention designs a second identification. When the first identification finds resonance, the second identification starts immediately. Fast Fourier Transform (FFT) is used to extract the I s_d and I s_q Amplitude-frequency characteristics. Through secondary resonance identification, the accuracy of resonance identification can be further ensured.

[0011] Fuzzy cognitive maps are used to construct a two-harmonic resonance identification model. As shown in Figure 3, since the dynamic iteration process of the fuzzy cognitive map model is highly consistent with the classification process of the time series, the fuzzy cognitive map shows a higher accuracy in time series classification compared with other advanced models. During training, the weight learning problem in the fuzzy cognitive map is transformed into a constrained convex optimization problem. The least squares method is introduced into this optimization problem to ensure the robustness and speed of the fuzzy cognitive map. This method can effectively improve the accuracy and real-time performance of resonance identification and provide more interpretable model results.

[0012] In the embodiment of the present invention, the inner current loop current is general and does not vary with the system voltage U s The current inner loop signal selected in the SVG controller of the SVG wind farm system includes: the current d-axis component I s_d , current q-axis component I s_q .

[0013] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention: At present, the wind farm high-frequency resonance identification method faces many problems, the main problems are: 1. At present, the high-frequency resonance analysis of wind farms is mostly studied from the operation of wind turbine converters. However, there are many wind turbines in the wind farm with different models, and the workload of electrical and control signal collection is large, which increases the difficulty of resonance analysis and identification.

[0014] 2. Previous high-frequency resonance identification models are often accompanied by defects such as low identification accuracy, slow identification speed, poor robustness, and difficulty in adapting to various big data identification situations.

[0015] Based on the above reasons, the present invention provides a method that can simply and accurately reflect the SVG outlet voltage Us and the current inner loop current (in Figure 2 Take the control mode in as an example) s_d ,I s_q The characteristic recognition model is put into the above resonance recognition application. To this end, the fuzzy cognitive map is used to avoid the limitations of the traditional resonance recognition model, such as low accuracy, slow recognition speed and poor robustness.

[0016] The present invention provides a method for online identification of high-frequency resonance of a wind farm system containing SVG based on fuzzy cognitive graph. The specific modeling steps are as follows: Figure 4 As shown, the method of the present invention comprises: Process S101, selecting the SVG outlet voltage and the current I in the inner current loop of its controller in the wind farm s_d ,I s_q .

[0017] In process S102, the SVG outlet voltage Us data set is converted into an interval form by granular computing in the first identification, and the current I in the current inner loop is extracted by fast Fourier transform in the second identification. s_d ,I s_q The amplitude-frequency characteristics.

[0018] In process S103, the interval data of each time window and the amplitude-frequency characteristics of each time window are fuzzy processed by fuzzy C-means clustering respectively; and the membership matrix of each time window is obtained.

[0019] In process S104, the obtained membership matrix is ​​labeled, 80% of the data is used as a training set, and after the fuzzy cognitive map resonance identification model is trained, 20% of the data is used as a test set and input into the resonance identification model for identification testing.

[0020] Process S105, embedding the trained resonance identifier into the SVG control system, such as Figure 5 As shown, if the first identification is normal, the system runs normally; otherwise, it is passed to the second identification, and the system makes corresponding responses according to the results of the second identification.

[0021] Combine the following Figure 6 The specific embodiments are as follows to introduce the technical solution of the present invention in detail: Step 1: In the first identification, the original time series of the SVG outlet voltage Us is converted into an interval time series by using the granular computing method. In the second identification, the current I in the controller current inner loop is transformed by using the fast Fourier transform. s_d ,I s_q The original time series is subjected to amplitude-frequency feature extraction.

[0022] Step 2: Use fuzzy C-means clustering to classify Us interval information particles or I s_d ,I s_q The amplitude-frequency characteristics of the data are fuzzified to calculate the degree of membership of each interval-type information particle or amplitude-frequency characteristic. This step uses fuzzy logic to handle the uncertainty and ambiguity of the data.

[0023] Step 3: Label each data point, and represent the resonant state and non-resonant state with numbers 1 and 2 respectively.

[0024] Step 4: Based on these labeled data, a resonance identification model based on fuzzy cognitive graph is established and trained. During the model training process, the resonant state and non-resonant state of the system are identified by calculating the membership degree of each information particle or amplitude frequency feature.

[0025] Step 5: After training is completed, the test set samples are brought into the trained model, the test samples are identified, and the identification results are quantitatively evaluated and analyzed.

[0026] Specifically, the SVG outlet voltage U s And the current I in the controller current inner loop s_d ,I s_q The waveform of the original data is shown in Figure 7. Each data set contains a non-resonance state, a strong resonance state, and a weak resonance state (a larger amplitude during resonance is called "strong resonance", and a smaller amplitude during resonance is called "weak resonance"). s And the current I in the controller current inner loops_d ,I s_q The time series X={x1,x2,x3,......,x t}, its length is t. The time series is divided into k windows with one cycle as the time window. The time length of each window is w. Then the relationship between k, t, and w is k = t / w. The divided time series is shown in formula 1:

[0027] During the first identification, granular computing is used to process data and the SVG outlet voltage U s The time series is divided into windows, and each time window is fuzzy processed using a triangular fuzzy set to obtain a fuzzy information granule G that can represent each window. Finally, a granular time series G = {G1, G2, ..., G K}, as shown in Formula 2. Among them, the granulation interval G K = {Low k ,R k ,Up k As shown in Figure 8, R is the average value of the change in the time window data, and Low and Up represent the lower and upper limits of the interval coverage amplitude respectively:

[0028] During the secondary identification, the fast Fourier transform is used for data processing to convert the current I s_d ,I s_q Each time window of is decomposed by fast Fourier transform, and the amplitude-frequency characteristic matrix F = {F1, F2, ..., F K} T As shown in Formula 3, where the amplitude F of each time window k ={F k1 ,F k2 ,...,F kw As shown in Figure 8:

[0029] The information particle G or amplitude-frequency characteristic F of each time window is regarded as a data sample. After clustering the samples by fuzzy C-means clustering, the membership matrix U = {U1, U2, ..., U K} T , where U k = {U k1 ,U k2 ,...,U Kc}, where c is the number of clusters of fuzzy C-means, and the membership matrix is ​​shown in Formula 4:

[0030] Fuzzy cognitive graphs of different sizes are established according to the number of clusters. The membership matrix U is used as learning data, and each learning data is labeled A = {A1,...,A K} T ,where A={A 11 ,...,A 1i}, i is the number of sample classification types. In this model, the main classifications are resonance (marked as 1) and non-resonance (marked as 2), so the value range of i is {1, 2}. The convex optimization algorithm is used to learn the fuzzy cognitive map. As shown in Formula 5: A i (t) = f(U(t)w i ) (5)

[0031] In formula (5), U(t)=[U1,U2,......,U c ], t=1,2,......,k is the state value of all nodes of FCM at time t, A i (t) is the system state value of the i-th node of FCM at time t, W i = {W 1i ,W 2i ,...,W ci} T is the i-th column of the weight matrix. Formula 5 can be transformed into Formula 6 through inverse transformation: f -1 (A i (t))=U(t)w i (6)

[0032] For each initial state vector, the system generates a response sequence containing k state values. These response sequences are organized in the form of matrices to eventually form a total of i state response matrices. According to Formula 6, the following linear equation can be generated as shown in Formula 7: Y i =UW i St|W ji ∣≤1,j=1,2,…,C (7)

[0033] In order to solve the weight, the following objective function is constructed:

[0034] Repeat the above steps to learn the weight matrix of all time series, and the fuzzy cognitive map completes the learning of the weight matrix W. Finally, the judgment of the resonant fault state value 1 or the non-resonant normal operation state value 2 is completed.

[0035] The accuracy of the system operation status obtained by model identification and the actual operation status is analyzed. The present invention uses the recognition accuracy of the fuzzy cognitive map system operation identifier to quantify the performance index analysis. The accuracy refers to the proportion of correctly identified samples to the total samples, and formula 9 is as follows: Among them, X T is the identification sample base; X error Identify the number of incorrect samples;

[0036] The identification error is shown in formula 10:

[0037] The first identification result of the model and the actual identification result as well as the identification error of the model are shown in Figure 9. The first resonance identification experiment was repeated 20 times, with an average accuracy of 98%. The identification result of the second identification model and the actual identification result as well as the identification error of the model are shown in Figure 9. The second resonance identification experiment was repeated 20 times, with an average accuracy of 99.5%. It can be seen from the identification speed that the first identification speed is much faster than the second identification. Therefore, the first identification ensures the speed of system resonance identification, and the second identification provides a guarantee for the accuracy of model identification.

[0038] In the embodiment of the present invention, after the resonance identification model is trained according to the parameters listed in Table 1, the identification logic is as follows: Figure 5 As shown in the figure, when the two identification models are started, the first identification model is entered first. When there is no resonance in the first identification model system, the system maintains the original mode operation; if resonance is identified, the model starts the second resonance identification to improve the identification accuracy of the system. The wind farm SVG simulation control system runs for 5 seconds, and resonance begins to appear in the third second. The identification results of the online resonance identifier are as follows Fig.10 As shown in the figure, although the online resonance identifier gives a wrong judgment at the initial startup, it is because in actual application, the resonance identification model monitors the system in real time to see if it is resonant when the system is stable. Therefore, the initial unstable state is not included in the identification result in the simulation. Fig.10 The system's resonance identification effect diagram shows that the system resonated at the third second. After 0.02 seconds of information collection, the system quickly made an identification at 3.02 seconds. After offline identification experiments and real-time identification experiments, it was found that the recognition rate of real-time identification was much faster than the offline identification time, and the resonance identification result was displayed almost immediately at the same time as the signal collection was completed.

[0039] Table 1 Table 2

Claims

1. A method for online identification of high-frequency resonance of a wind field containing SVG based on fuzzy cognitive graph, characterized in that: The steps are: S1, SVG outlet voltage U s And the current I in the controller current inner loop s_d ,I s_q The time series X={x1,x2,x3,......,x t } indicates that its length is t. The time series is divided into k windows with one cycle as the time window. The time length of each window is w. Then the relationship between k, t, and w is k = t / w. The time series after segmentation is: During the first identification, the SVG output voltage U s The time series is divided into windows, and each time window is fuzzy processed using a triangular fuzzy set to obtain a fuzzy information granule G representing each window, and finally a granular time series G = {G1, G2, ..., G K }, as shown in Formula 2: Among them, the granulation interval G K = {Low k ,R k ,Up k }, R is the average value of the change in the time window data, Low and Up represent the lower and upper limits of the interval coverage amplitude respectively; During the secondary identification, the current I in the inner loop of the controller is s_d ,I s_q Each time window of is decomposed by fast Fourier transform to obtain the amplitude-frequency characteristic matrix F = {F1, F2, ..., F K } T , as shown in Formula 3: Among them, the amplitude F of each time window k ={F k1 ,F k2 ,...,F kw }; The information particle G or the amplitude-frequency characteristic F of each time window is regarded as a data sample. After clustering the samples by fuzzy C-means clustering, the membership matrix U = {U1,U2,...,U K } T , the membership matrix is ​​shown in Formula 4: Among them U k = {U k1 ,U k2 ,...,U Kc }, c is the number of clusters of fuzzy C-means; Fuzzy cognitive graphs of different sizes are established according to the number of clusters. The membership matrix U is used as learning data, and each learning data is labeled A = {A1,...,A K } T , where A={A 11 ,...,A 1i }, i is the number of sample classification categories, and the convex optimization algorithm is used to learn the fuzzy cognitive map, as shown in Formula 5: Ai(t)=f(U(t)wi)(5) Where U(t)=[U1,U2,...,U c ], t=1,2,......,k is the state value of all nodes of FCM at time t, A i (t) is the system state value of the i-th node of FCM at time t, W i = {W 1i ,W 2i ,...,W ci } T is the i-th column of the weight matrix; Formula 5 can be transformed by inverse transformation: f-1(Ai(t))=U(t)wi(6) According to formula 6, the following linear equation can be generated: Yi=UWi S.t.∣W ji ∣≤1,j=1,2,…,C(7) Objective function: min||U*wi-Yi||2+β||wi||1 st ||wi||∞≤1 (8) Repeat the above steps to learn the weight matrix of all time series, and the fuzzy cognitive map completes the learning of the weight matrix W; S2, the recognition accuracy of the identifier running with the fuzzy cognitive map system: Among them, X T is the identification sample base; X error Identify the number of incorrect samples; The identification error is:

Images