Double-layer identification method for subsynchronous oscillation generated by series compensation grid connection of fan group

Abstract

The invention belongs to the related technical field of wind power generation, and particularly relates to a double-layer identification method for subsynchronous oscillation generated by series compensation grid connection of a fan group, which comprises the following steps of: acquiring power quality parameter data of a grid connection point and each generator set in a target fan group, and constructing a one-dimensional power quality parameter matrix; a data image conversion function is adopted to convert the parameter matrix into an RGB color block map representing oscillation characteristics; inputting the RGB color block map into an oscillation level identification model to obtain an oscillation level; inputting the RGB color block map with the oscillation level being a preset level type into an oscillation generation position identification model to obtain an oscillation generation position, and completing double-layer identification of the subsynchronous oscillation generated by the target fan group through series compensation grid connection; the electric energy quality parameter matrix sequentially comprises voltage, current, frequency, active power, reactive power, a voltage harmonic value, a current harmonic value and a rotor rotating speed. According to the invention, instant and generalization identification of subsynchronous oscillation can be realized when an actual fault occurs.

Description

Technical Field

[0001] The present invention belongs to the technical field related to wind power generation, and more specifically, relates to a two-layer identification method for subsynchronous oscillation generated by a wind turbine group connected to the grid through series compensation. Background Art

[0002] Subsynchronous oscillation (SSO) refers to an oscillation phenomenon in a power system where the frequency is lower than the system base frequency (usually 50 Hz or 60 Hz). It is an unstable oscillation caused by the interaction between mechanical, control, and electrical factors within the system. Subsynchronous oscillation (SSO) is a common oscillation problem in power systems. It was initially mainly concentrated in large steam turbine generator sets, and the torsional vibration problem of the shafting caused by its shafting structural characteristics has attracted extensive attention. With the rapid development of the wind power industry and the increasing proportion of renewable energy generation, the penetration rate of wind turbines in the power grid is continuously increasing, and the SSO problem has gradually become a major hidden danger in wind power systems. The dynamic characteristics of wind turbines are different from those of traditional steam turbine generators. Their shafting stiffness is relatively low and the rotor inertia is relatively large, resulting in a torsional vibration mode frequency lower than 3 Hz and weak damping. However, due to the relatively low torsional vibration mode frequency of wind turbine units, it is usually difficult to directly interact with the resonance mode of the power grid. Therefore, in wind farms, the induction generator effect (IGE) and subsynchronous control interaction (SSCI) have become the main factors causing SSO, rather than the traditional shafting torsional vibration. Especially in the scenario of using series compensation, the resonance introduced by series capacitors will amplify the subsynchronous oscillation in the electrical system, seriously threatening the stability and power generation efficiency of wind farms.

[0003] To improve the efficiency of long-distance power transmission, series compensation devices are often used in offshore and onshore wind farms to connect to the grid. However, although series compensation can improve the power transmission capacity, it also causes more serious and diverse subsynchronous oscillation problems. Existing research shows that in a wind farm based on a doubly-fed induction generator (DFIG) connected to the grid through series compensation, the frequency of the subsynchronous oscillation current is approximately 20 Hz, and its amplitude and damping characteristics are closely related to the wind speed, series compensation degree, control strategy, and parameters of the grid-side converter (GSC). When the wind power output increases, the amplitude of the SSO current also rises, which may lead to serious system instability and increase the risk of wind curtailment. In addition, the influence degree and spread degree of different levels of SSO are different; for high-risk SSO, it is necessary to quickly complete identification and locate its generation location. In response to this challenge, it is crucial to develop an effective subsynchronous oscillation generalization and real-time identification strategy. Summary of the Invention

[0004] In view of the above defects or improvement requirements of the prior art, the present invention provides a two-layer identification method for subsynchronous oscillation generated by a wind turbine group connected to the grid through series compensation, aiming to solve the problem that in the existing research on subsynchronous oscillation generated by wind turbines integrating into the large grid, most of the research focuses on mechanism analysis, and there is a lack of immediate and generalized identification strategies for subsynchronous oscillation when actual faults occur.

[0005] To achieve the above object, according to one aspect of the present invention, there is provided a two-layer identification method for subsynchronous oscillation generated by a wind turbine group connected to the grid through series compensation, including:

[0006] Collect the power quality parameter data of the connection point and each generating unit in the target wind turbine group, and construct a one-dimensional power quality parameter matrix; use a data image conversion function to convert the one-dimensional power quality parameter matrix into an RGB color block diagram representing oscillation characteristics;

[0007] Input the RGB color block diagram into the oscillation level identification model to obtain the oscillation level; determine whether the oscillation level is a preset level type, if so, input the RGB color block diagram into the oscillation generation position identification model to obtain the oscillation generation position, and complete the two-layer identification of the subsynchronous oscillation generated by the target wind turbine group connected to the grid through series compensation;

[0008] Among them, the one-dimensional power quality parameter matrix sequentially includes voltage, current, frequency, active power, reactive power, voltage harmonic value, current harmonic value, and rotor speed.

[0009] Further, both the oscillation level identification model and the oscillation position identification model are trained using a neural network structure.

[0010] Further, both the oscillation level identification model and the oscillation position identification model are trained using a deep convolutional neural network.

[0011] Further, the Matlab software Image function is used to implement data image conversion.

[0012] Further, the output of the oscillation level identification model includes three level types: decaying oscillation, equal-amplitude oscillation, and divergent oscillation.

[0013] Further, the preset level type is equal-amplitude oscillation and / or divergent oscillation.

[0014] According to another aspect of the present invention, there is provided an electronic device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0015] According to another aspect of the present invention, there is provided a computer-readable storage medium, which includes a stored computer program. When the computer program is run by a processor, it controls the device where the storage medium is located to execute the steps of the method described above.

[0016] According to another aspect of the present invention, there is provided a computer program product, including a computer program or instruction. When the computer program or instruction is executed by a processor, it realizes the steps of the method described above.

[0017] Generally speaking, compared with the prior art by the above technical solution conceived by the present invention, the technical solution provided by the present invention mainly has the following beneficial effects:

[0018] 1. The method of the present invention is based on the RGB color block diagram characterizing the oscillation characteristics. Through the constructed oscillation level recognition model and oscillation position recognition model, it realizes the dual recognition of the oscillation level and the generation position corresponding to a specific level. Since this method is based on AI technology, it can achieve fast recognition and has high generalization ability, effectively avoiding the problem of poor generalization brought by the traditional method relying on physical models for recognition. At the same time, this method can set the target level and, according to needs, recognize the generation position of a specific oscillation level, with high flexibility and strong practicability, and does not need to obtain the generation position based on all data without distinction like the traditional method. In addition, the RGB color block diagram adopted by this method is a one-dimensional power quality parameter matrix based on specific elements and sorting, and the obtained RGB color block diagram has a high degree of discrimination for different oscillation levels, ensuring the recognition accuracy. Therefore, the method of the present invention can realize the instant and generalized recognition strategy of subsynchronous oscillation when an actual fault occurs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of a double-layer recognition method for subsynchronous oscillation generated by a wind turbine group connected to the grid through series compensation provided by an embodiment of the present invention;

[0020] Figure 2 is a system topology diagram of four groups of doubly-fed wind turbine generators connected to the grid through series compensation equipment provided by an embodiment of the present invention;

[0021] Figure 3 is a schematic diagram of the decay oscillation type provided by an embodiment of the present invention;

[0022] Figure 4 is a schematic diagram of the equal-amplitude oscillation type provided by an embodiment of the present invention;

[0023] Figure 5 is a schematic diagram of the divergence type provided by an embodiment of the present invention;

[0024] Figure 6It is a schematic diagram of the data matrix provided by the embodiments of the present invention;

[0025] Figure 7 It is an RGB color block diagram corresponding to the damped oscillation provided by the embodiments of the present invention;

[0026] Figure 8 It is an RGB color block diagram corresponding to the equal-amplitude oscillation provided by the embodiments of the present invention;

[0027] Figure 9 It is an RGB color block diagram corresponding to three level types of divergent oscillations provided by the embodiments of the present invention;

[0028] Figure 10 It is a schematic diagram of the mechanism of the conventional neural network provided by the embodiments of the present invention;

[0029] Figure 11 It is a schematic diagram of the specific training effect obtained by the FNN method provided by the embodiments of the present invention;

[0030] Figure 12 It is a schematic diagram of the reason for choosing Capsnet provided by the embodiments of the present invention;

[0031] Figure 13 It is a schematic diagram of the change in the training accuracy of the capsule neural network provided by the embodiments of the present invention;

[0032] Figure 14 It is a schematic diagram of using max pooling in the pooling layer provided by the embodiments of the present invention;

[0033] Figure 15 It is a schematic diagram of the training accuracy of the deep convolutional neural network provided by the embodiments of the present invention;

[0034] Figure 16 It is a schematic diagram of generating a position recognition result provided by the embodiments of the present invention;

[0035] Figure 17 It is a schematic diagram of the overall training process and accuracy rate of the double-layer model provided by the embodiments of the present invention;

[0036] Figure 18 It is a flowchart of the overall implementation process provided by the embodiments of the present invention. Specific implementation manners

[0037] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0038] Embodiment 1

[0039] A two - layer identification method for the subsynchronous oscillation generated by a wind turbine group connected to the grid through series compensation, as Figure 1 shown, includes:

[0040] Collect the power quality parameter data of the grid connection point and each generating unit in the target wind turbine group, and construct a one - dimensional power quality parameter matrix; use a data image conversion function to convert the one - dimensional power quality parameter matrix into an RGB color block diagram representing oscillation characteristics;

[0041] Input the RGB color block diagram into the oscillation level identification model to obtain the oscillation level; judge whether the oscillation level is a preset level type. If so, input the RGB color block diagram into the oscillation generation position identification model to obtain the oscillation generation position, and complete the two - layer identification of the subsynchronous oscillation generated by the target wind turbine group connected to the grid through series compensation;

[0042] Among them, the one - dimensional power quality parameter matrix sequentially includes voltage, current, frequency, active power, reactive power, voltage harmonic value, current harmonic value, and rotor speed.

[0043] The method of this embodiment is based on the RGB color block diagram representing oscillation characteristics, and through the constructed oscillation level identification model and oscillation position identification model, it realizes the identification of the oscillation level and the dual identification of the generation position corresponding to a specific level. Since this method is based on AI technology, it can achieve fast identification, and has high generalization ability, effectively avoiding the problem of poor generalization brought by the traditional method relying on physical models for identification. At the same time, this method can set the target level, and according to needs, identify the generation position of a specific oscillation level, with high flexibility and strong practicability, without the need to obtain the generation position based on all data without distinction like the traditional method. In addition, the RGB color block diagram adopted by this method is based on a one - dimensional power quality parameter matrix with specific elements and sorting, and the obtained RGB color block diagram has a high discrimination degree for different oscillation levels, ensuring the identification accuracy. Therefore, the method of this embodiment can realize the real - time and generalization identification strategy of subsynchronous oscillation when an actual fault occurs.

[0044] As a preferred implementation, both the oscillation level identification model and the oscillation position identification model are trained using a neural network structure.

[0045] As a preferred implementation, both the oscillation level identification model and the oscillation position identification model are trained using a deep convolutional neural network.

[0046] As a preferred implementation, the Matlab software Image function is used to implement data image conversion.

[0047] As a preferred embodiment, the output of the oscillation level recognition model includes three level types: decaying oscillation, equal-amplitude oscillation, and divergent oscillation.

[0048] As a preferred embodiment, the preset level types are equal-amplitude oscillation and / or divergent oscillation.

[0049] Common wind power generation generally uses a doubly-fed induction generator (DFIG). Therefore, taking the DFIG as an example, the construction of the sub-synchronous oscillation topology model will be described.

[0050] The DFIG is connected to the power grid through a filter and a rotor-side converter (RSC), enabling the wind power generation system to achieve variable-speed constant-frequency control. The use of series compensation capacitor technology can effectively reduce transmission line losses, increase the line transmission capacity, increase the stability margin of the system, and its cost is much lower than that of high-voltage direct current transmission (HVDC) and controlled series compensation technology.

[0051] The series compensation k C level of the DFIG grid-connected power generation system is defined as follows:

[0052] k C = X C / X LΣ

[0053] In the formula, X C is the reactance of the series compensation capacitor, and X LΣ is the total reactance of the line.

[0054] For a single wind turbine, factors such as the series compensation capacitor value, input wind speed, and internal loop parameters controlled by the rotor-side converter (RSC) will affect the stability of the system and may lead to divergent oscillations. However, considering the actual operating environment of a wind farm, the sub-synchronous oscillations caused by a single wind turbine are usually limited. Therefore, this study mainly explores the sub-synchronous oscillation problems of four groups of doubly-fed wind turbine generators under coupled conditions. The system topology is as Figure 2 shown.

[0055] For a single DFIG system, its stator is directly connected to the power grid, and the rotor realizes AC excitation through a three-phase AC-DC-AC converter. The DFIG realizes power exchange through a dual-channel between the stator, rotor, and the power grid. The rotor-side converter uses a three-phase two-level voltage-source PWM converter with a back-to-back configuration. According to their positions, these two groups of PWM converters are respectively called the grid-side converter (GSC) and the rotor-side converter (RSC), and they are connected through a DC bus capacitor.

[0056] The GSC is connected to the power grid, converting alternating current into direct current to provide a stable DC voltage for the DC bus capacitor; the RSC converts direct current into alternating current, regulating the excitation current frequency and amplitude of the DFIG rotor, thereby achieving speed control and the regulation of active and reactive power. The RSC control system realizes the precise control of power electronic devices through a PWM inverter, thereby achieving maximum wind energy tracking and the safe, stable, and efficient grid connection operation of the wind turbine. This control design ensures the accurate reproduction of the actual operating conditions of the wind turbine by the simulation system, enabling the system to operate stably under changing wind speeds and maintain good power exchange with the power grid.

[0057] Regarding the SSO classification, the description is as follows:

[0058] (1) Classify the SSO level according to the impact degree

[0059] According to different oscillation intensities and diffusion capabilities, SSO can be divided into three different categories - damped oscillation (as shown in Figure 3 ), equal-amplitude oscillation (as shown in Figure 4 ), and divergent oscillation (as shown in Figure 5 ).

[0060] Among them, as shown in the figure, the system is in an unstable grid connection stage within the first two seconds, and the oscillation generated from 2 to 5 seconds is sub-synchronous oscillation. Therefore, during the process of data extraction and processing, only the operation data from 2 to 5 s is selected.

[0061] Taking the waveform diagram of the active power at the grid connection point as an example, it can be seen that the damped oscillation has a relatively small impact on the system and will dissipate autonomously over time; the equal-amplitude oscillation has the second-largest amplitude, is difficult to self-dissipate, but also does not spread significantly, and its impact on the overall power grid is between the damped oscillation and the divergent oscillation; the divergent oscillation has a large amplitude and will intensify over time, with the characteristics of large impact and fast diffusion speed. If not discovered and measures taken in time, it is extremely likely to cause large-scale oscillation or even splitting of the system. Therefore, the rapid, real-time, and accurate determination of the SSO level, and the rapid positioning of the wind farm where the oscillation occurs based on the power quality parameters at the grid connection point during equal-amplitude oscillation and divergent oscillation, play an important role in the targeted practical treatment and suppression of SSO.

[0062] (2) Classify the SSO generation location according to the wind farm number

[0063] Most of the existing wind power data acquisitions are completed at the grid connection point. Therefore, when SSO occurs, judging the location of the wind turbine where the oscillation occurs based on the grid connection point data has important practical application value for quickly and accurately taking corresponding measures to suppress SSO or cut off the wind farm with severe oscillation to avoid large-scale oscillation diffusion.

[0064] Regarding the construction of the one-dimensional power quality parameter matrix, the description is as follows:

[0065] Simulations were carried out for nine typical operating conditions, including: wind speeds v = 5 m / s, 7 m / s, 10 m / s, under reactive power compensation conditions of 15%, 35%, and 65% respectively. The converter parameters were changed, and parameters such as voltage, current, frequency, active power, reactive power, harmonic values of voltage and current, and rotor speed of the wind turbine group and the intersection point were collected, and databases were established according to the above two classification methods. Each group of data was constructed into a data matrix (as Figure 6 shown). The structure of the data matrix was designed to generate more obvious image color differences when subsynchronous oscillation occurred, so as to improve the recognition effect of the neural network.

[0066] In order to achieve the local unified features and global difference features of data in the image, it is necessary to spatially separate the power quality parameter group from the wind turbine speed data. By aggregating similar physical quantities, the same physical quantities between the summary point and different wind turbines are placed in the same row, and the relatively similar physical quantities in the same unit are placed in the same column. The rows of the obtained matrix (as Figure 6 shown) are spliced end to end to obtain the power quality parameter vector, which is the above-mentioned one-dimensional power quality parameter matrix and conforms to the pooling principle of the neural network pooling layer.

[0067] The RGB color block diagrams corresponding to the three level types of decaying oscillation, equal-amplitude oscillation, and divergent oscillation are as Figure 7 , Figure 8 , Figure 9 shown.

[0068] Regarding the selection of the recognition network structure comparison, the following is explained:

[0069] After clarifying the basic situation classification and data characteristics of the wind turbine subsynchronous oscillation, it is necessary to further perform SSO intelligent recognition based on the converted results of the existing data. For the autonomous recognition process faced by the present invention, based on methods such as neural networks, the design of intelligent agents is the basic research direction.

[0070] (1) Conventional neural network

[0071] Among the common intelligent agent design methods, the conventional neural network method is the most easily considered research method. By converting the wind turbine operation data, it can be adapted to the data processing requirements of the conventional neural network. After inputting the feature data, it is trained with the training data to make the parameters of the neural network fit the actual determination requirements. Finally, the "black box" of the neural network can intelligently complete the determination of the wind turbine subsynchronous oscillation.

[0072] Specifically, as Figure 10As shown in the figure, the state characteristic pictures of the fan need to be reconstructed first to form a 128×128 RGB three-dimensional tensor. Then, the tensor is transformed into a one-dimensional tensor that can be processed by the neural network through flatten. Then, the training and fitting of the black-box model are completed through the feedforward neural network architecture (FNN). Finally, the trained neural network agent can output a three-dimensional state vector to describe the three different sub-synchronous oscillation states of the wind turbine.

[0073] After training and fitting, the FNN method has achieved the specific training effect as Figure 11 shown. In terms of the convergence process of the training effect, the FNN method can basically effectively determine the sub-synchronous state of the fan, and the accuracy rate can reach 86%. Further analyzing the confusion matrix of the model, it can be clearly observed that the FNN method has the best judgment effect on decaying oscillations and also has a good prediction for divergent oscillations. However, limited by the relatively complex picture characteristics of the equal-amplitude oscillation state, the method has a poor judgment on the equal-amplitude oscillation.

[0074] From the specific results, as shown in Table 1 below, the determination of the equal-amplitude oscillation state itself is often relatively accurate. However, different equal-amplitude oscillation states often cannot derive consistent results, which leads to the FNN method missing a considerable number of equal-amplitude oscillation states, that is, the recall is not ideal. In addition, the misjudgment of the divergent oscillation is also relatively serious, which also makes the determination balance (f1-score) of the divergent oscillation state poor. Therefore, it needs to be solved through further algorithm optimization.

[0075] Table 1 Recognition Results

[0076]

[0077] (2) Capsule Neural Network

[0078] The Capsule Neural Network (CapsNet) analyzes the overall information by aggregating sequence information. At present, the Capsule Neural Network has been widely used in medical image analysis and, as a graph analysis network, captures global information through local information.

[0079] CapsNet consists of a convolutional layer, a Primary Capsule Layer, and a DigitCapsule Layer. Among them, the capsule layer acts as a feature extractor of the network, mapping low-dimensional features to high-dimensional features. Features such as texture and color in the image are stored in capsules composed of multiple neurons. Each capsule can predict the global features of the overall entity based on the attributes of its internal neurons.

[0080] The input layer and fully connected layer of CapsNet are similar to those of traditional convolutional neural networks, but a coupling coefficient is introduced in the linear summation stage. The input S of the capsule network can be obtained through the following formula:

[0081]

[0082] where u i is the output of the previous layer of the capsule network, W ij is the weight matrix multiplied by each output, and C ij is the coupling coefficient, representing the connection strength between each capsule neuron in the previous layer and a specific neuron in the next layer.

[0083] The output v of the capsule network is calculated through the following formula:

[0084]

[0085] where v j is the network output, and s j is the network input. The Squash activation function is used for non-linear processing to compress the amplitude of the output vector v j to the range of 0 to 1. The amplitude of the vector v j represents the probability of the corresponding category, and the vector itself contains information about features, positions, and directions.

[0086] Through this unique feature aggregation and coupling mechanism, the capsule network can effectively capture and represent the spatial relationships in the data and is suitable for the analysis of complex systems. In this study, CapsNet is used to analyze the sub-synchronous oscillation modes of a wind turbine group, and its superiority in capturing the dynamic characteristics and global information of the wind power system is verified through comparison with other deep learning networks.

[0087] To further optimize the oscillation state determination problem, specific problem analysis and algorithm optimization are required. Different from the stable state, the oscillation state produces a color difference shift in the result combination elements, as shown in Figure 12 . However, for the converted results of pictures belonging to the oscillation state, there is a significant relative shift effect of the combination elements as a whole. This matches the application scenario of the conventional capsule neural network applied to spatial effect conversion.

[0088] Collect real-time state variables under different wind speeds, numbers of wind turbines, converter parameters, and series compensation degrees, obtain a data matrix with a fixed sampling period, and perform preprocessing to generate data pictures for network training and verification. Based on the existing data scale, a data structure with multiple epoch iterations is formed, and 200 epoch cycles are used for training. Finally, the specific training process and the change of training accuracy are shown in Figure 13 .

[0089] In terms of the specific training effect, CapsNet can achieve an identification accuracy of up to 91%, which is a significant improvement compared to the conventional neural network method described above. From the perspective of the change process of accuracy and loss value, it shows a gradually convergent training characteristic, which is consistent with the general training process of neural networks.

[0090] It is worth noting that in the later epochs of training, CapsNet exhibits relatively obvious fluctuations. At the same time, the overall accuracy is also slowly increasing. This indicates that under the current data support conditions, CapsNet is difficult to be fully trained to achieve the best results. This is mainly because the training process of CapsNet relies on a dynamic routing mechanism, and the optimization of this mechanism is relatively complex and difficult to converge to the optimal value as quickly as traditional networks. Therefore, more effective structural optimization is needed.

[0091] (3) Deep Convolutional Neural Network (DCNN)

[0092] Compared with the capsule neural network, DCNN has more prominent features. With sufficient computing power, a deep convolutional neural network with deeper hidden layers and a more complex network structure can perform more complex information processing tasks. In terms of performance, the "black box" model composed of a deep neural network can achieve a more complex analysis and fitting effect compared to the capsule neural network. For the SSO analysis of grid-connected wind turbines, the enhanced analysis and fitting ability means a more comprehensive analysis of complex wind farms, thus effectively reducing misjudgments in SSO detection. This is also an important reason why a deep convolutional neural network is mainly used for SSO analysis in this embodiment.

[0093] A deep convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer of this network receives a large amount of data from a wind farm (n wind turbines). After the data collection is completed in the input layer, the data will be preprocessed to form a two-dimensional data field. The first group of convolutional layers is used to identify low-level features of the image (such as edges). After the feature extraction is completed, the pooling layer screens and filters the information. Then, through convolutional and pooling operations, the next group of convolutional layers gradually extracts more complex features based on the low-level features. Finally, the network enters the fully connected layer (i.e., the classifier). In the fully connected layer, all the feature matrices of the pooling layer are converted into one-dimensional feature vectors to reduce the influence of feature positions on the classification results.

[0094] After the data is input through the input layer, the convolutional layer calculates the dot product between a specific region of the input image and the weight matrix (i.e., the filter). This filter slides across the entire image, and the dot product operation is repeated to extract image features. The number of channels of the filter must be the same as that of the input image. Generally, as the network depth increases, more filters are used, which means more edges and features can be detected. From the first pixel to the last pixel of the image, each pixel is convolved with the filter to extract features from the image and generate a convolutional layer representing the image features.

[0095] After the convolution operation is completed, a pooling operation is performed to reduce the vector dimension of the output data of the convolutional layer, reduce the computational complexity, and prevent overfitting. Although the image resolution is reduced, the pooling layer retains the effective information of various features of the image and reduces the processing complexity of each layer of the network, thereby improving the running speed of the network. Common pooling methods include max pooling and average pooling. The network uses max pooling as the pooling method. Figure 14 Shows the pooling calculation with a stride of 2, where max(7, 2, 5, 6) = 7. In addition, overlapping pooling and spatial pyramid pooling also show good performance in some applications of convolutional neural networks.

[0096] The purpose of the fully connected layer is to map the features extracted by convolution and pooling to the label space of the samples to achieve classification and output control signals. The fully connected layer converts the two-dimensional feature map generated by the convolutional layer and the pooling layer into a one-dimensional vector. In the fully connected layer, the input nodes of each layer are connected to all the nodes of the next layer, which means that each neuron in the fully connected layer is related to all the neurons of the previous layer. A fully connected network with sufficient depth provides a structural basis for the complex analysis of deep neural networks and is crucial in SSO analysis.

[0097] Based on the above comparative analysis, under different wind speeds, the number of wind turbines, converter parameters, and compensation degrees, real-time state variables are collected, and the data is sampled, preprocessed at a fixed sampling period, and converted into data images for network training and verification. 70% of the data images are used for training, and 30% of the data images are used for verification. This embodiment uses a deep convolutional neural network with 144 layers, and its basic structure is shown in the following figure. The learning rate is set to 0.001, and this neural network is trained.

[0098] It should be noted that it is difficult for a deep convolutional neural network without specific optimization to achieve good training convergence. Therefore, it is necessary to optimize the selection of the network and parameters in combination with the requirements of the SSO (sub-synchronous oscillation) analysis environment.

[0099] 1) It is necessary to select an appropriate number of iterations. During the SSO analysis process, the characteristics of wind turbines are highly complex, and there are many interfering characteristics. If the number of iterations is too large, it may lead to a deviation in feature focus; conversely, if the number of iterations is too small, the convolutional neural network cannot complete sufficient learning on a large-scale complex data set. Therefore, an appropriate number of iterations must be selected. In this example, 214 iterations are performed for 5 rounds.

[0100] 2) It is necessary to consider the data selection range. For example, when the wind speed is 5 m / s, kc = 40%, kp = 3, ki = 0.6, and the line resistance is 0.08 Ω, the first 40,412 data samples are all 0. Since these data cannot fully represent the characteristics under this working condition, this part of the data should be excluded when selecting training data to reduce the convergence difficulty caused by incorrect data sources.

[0101] The training accuracy results are as Figure 15 shown. For the sub-synchronous oscillation level recognition network based on CNN, the total amount of training and validation data used in each group is 200, and a recognition accuracy of 100% is finally achieved.

[0102] It should be noted that the accuracy has been relatively high since the initial stage of training (after one round of iteration), which proves the potential of using DCNN for real-time recognition of SSO.

[0103] For the above statements and arguments, DCNN is selected to construct a two-layer SSO recognition method. The first layer of this method realizes the recognition and determination of the sub-synchronous oscillation level, which is beneficial for judging whether measures need to be taken in practical applications; the second layer realizes the positioning of the location where the sub-synchronous oscillation occurs, that is, among the sub-synchronous oscillations generated by the grid connection of multiple wind farms, through the analysis of the power quality at the grid connection point, it is possible to identify which wind farm the wind turbine generating the sub-synchronous oscillation is located in.

[0104] The recognition accuracy of the first layer is 100% when the data set is 1000 data per group and the number of iterations is five rounds. However, due to the rapid spread and great harm of SSO, real-time recognition under a small data set is also important.

[0105] Table 2 Prediction accuracy of different neural networks

[0106]

[0107] As can be seen from the above table, the first-layer recognition method constructed by DCNN performs well under both large and small data sets, and can achieve instant and accurate recognition of the SSO level.

[0108] The example of the recognition result map of the wind farm where the second-layer oscillating wind turbine is located is as Figure 16 shown. The overall training process and accuracy are as Figure 17As shown, the recognition accuracy of the second layer reaches 99.7%. The overall implementation process of this embodiment is as follows Figure 18 shown.

[0109] Embodiment 2

[0110] This application also relates to an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0111] The electronic device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The so-called processor can be a Central Processing Unit (CPU), or can also be other general-purpose processors, Digital Signal Processors (DSPs), Application-Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory, various functions of the electronic device can be realized.

[0112] The related technical solutions are the same as above and will not be elaborated here.

[0113] Embodiment 3

[0114] This application also relates to a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0115] Specifically, the memory can include high-speed random access memory, and can also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0116] The related technical solutions are the same as above and will not be elaborated here.

[0117] Embodiment 4

[0118] An embodiment of the present application provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps of the method in the above embodiment of the present application.

[0119] The related technical solutions are the same as above and will not be elaborated here.

[0120] Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A two-layer identification method for subsynchronous oscillations generated by wind turbines connected to the grid through series compensation, characterized in that: include: Collect the power quality parameter data of the grid connection point and each generator set in the target wind turbine group and construct a one-dimensional power quality parameter matrix; Using a data image conversion function, the one-dimensional power quality parameter matrix is ​​converted into an RGB color block image representing oscillation characteristics; Input the RGB color block image into the oscillation level recognition model to obtain the oscillation level; determine whether the oscillation level is a preset level type, and if so, input the RGB color block image into the oscillation generation position recognition model to obtain the oscillation generation position, thereby completing the double-layer recognition of the sub-synchronous oscillation generated by the target wind turbine group after series compensation and grid connection; The one-dimensional power quality parameter matrix sequentially includes voltage, current, frequency, active power, reactive power, voltage harmonic value, current harmonic value and rotor speed.

2. The double-layer recognition method according to claim 1, characterized in that: The oscillation level recognition model and the oscillation position recognition model are both obtained by training with a neural network structure.

3. The double-layer recognition method according to claim 2, characterized in that: The oscillation level recognition model and the oscillation position recognition model are both obtained by training with a deep convolutional neural network.

4. The double-layer recognition method according to claim 1, characterized in that: The image conversion is realized by using the Image function of Matlab software.

5. The double-layer recognition method according to claim 1, characterized in that: The output of the oscillation level identification model includes three level types: decaying oscillation, constant amplitude oscillation and divergent oscillation.

6. The double-layer recognition method according to claim 5, characterized in that: The preset level type is constant amplitude oscillation and / or divergent oscillation.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed by a processor, the device where the storage medium is located is controlled to execute the steps of the method according to any one of claims 1 to 6.

9. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images