Oscillating source field station single field station positioning removal method based on meta learning, electronic equipment and storage medium

Abstract

The invention discloses an oscillating source station single-station positioning and removing method based on meta-learning, electronic equipment and a storage medium, and relates to the technical field of power systems and automation thereof. Comprising the following steps: acquiring oscillation operation data of a single station, the oscillation operation data comprising electrical characteristics of a station exit; performing feature extraction on the oscillation operation data through a pre-trained neural network model, and judging whether the single station is an oscillation source or not; generating an optimal initial parameter of the neural network model based on a meta-learning process, wherein the meta-learning process is trained by using a forced oscillation sample; and if the single-field station is an oscillation source, triggering a circuit breaker to execute cutting operation, and suppressing oscillation propagation. Compared with a traditional oscillation source positioning method, the method avoids a data transmission process between stations, avoids communication difficulty, and has engineering practicability and timeliness.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, specifically to a method for locating and removing oscillation source stations in a single station based on meta-learning, an electronic device, and a storage medium. Background Technology

[0002] As the proportion of new energy power generation, represented by direct-drive wind power, continues to increase, the system's dimensionality is constantly rising, multi-timescale problems are becoming more prominent, the dynamic interaction process of various system components is becoming more complex, and subsynchronous / supersynchronous oscillations are occurring frequently, posing a severe challenge to the stable operation of the power system. For existing projects, disconnecting the oscillation source station is the most direct and feasible method for oscillation suppression. Disconnecting the oscillation source station first requires precise location of the oscillation. In order to suppress oscillation propagation more promptly, research is urgently needed on methods for rapid and accurate oscillation source location and disconnection.

[0003] Currently, there are two main types of oscillation source localization techniques: traditional physical methods and machine learning methods. Traditional physical methods mainly include energy flow methods, modal analysis methods, and hybrid dynamic simulation methods. However, each of these methods has its limitations. Energy flow methods and modal analysis methods require accurate modeling, but as the complexity of power systems continues to increase, the mathematical models required for these methods are becoming increasingly difficult to establish accurately. Hybrid dynamic simulation methods require multiple simulations, which takes a long time and cannot meet the timeliness requirements for oscillation localization. Machine learning-based oscillation source localization algorithms avoid the complex modeling process and can respond quickly, effectively solving the modeling difficulties and timeliness problems of traditional physical methods. However, in practical applications, these algorithms need to transmit all data to a unified collection station for analysis, which presents challenges in data transmission. Considering the time required for data transmission from the power station to the collection station, and the time required for the collection station to transmit power station disconnection commands to the power station, these methods cannot meet the timeliness requirements for rapid oscillation suppression. Furthermore, during wireless communication transmission, signals are prone to noise interference and data loss, testing the robustness of the algorithms.

[0004] Therefore, there is an urgent need to study a method for fast and accurate localization of machine learning oscillation sources to solve the problem of data transmission difficulties. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of difficult data transmission in the prior art, and to provide a single-site location and removal method for oscillation source stations based on meta-learning, an electronic device, and a storage medium.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A single-site location and removal method for oscillation source stations based on meta-learning includes the following steps:

[0008] Acquire oscillation operation data for a single station, the oscillation operation data including the electrical characteristics of the station outlet;

[0009] The oscillation operation data is used to extract features by a pre-trained neural network model to determine whether the single station is an oscillation source;

[0010] The optimal initial parameters of the neural network model are generated based on a meta-learning process, which is trained using forced oscillating samples.

[0011] If the single station is an oscillation source, the circuit breaker is triggered to perform a disconnection operation to suppress the oscillation propagation.

[0012] As a preferred technical solution:

[0013] As described above, the single-site location and removal method for oscillation source stations based on meta-learning includes acquiring the oscillation operation data of a single station, which comprises:

[0014] The measuring device of the single station collects the electrical characteristics of the station outlet, including active power, reactive power, phase A voltage and phase A current;

[0015] The electrical features are subjected to max-min normalization to generate normalized input features.

[0016] The single-site location and removal method for oscillation source stations based on meta-learning, as described above, involves feature extraction of the oscillation operation data using a pre-trained neural network model, including:

[0017] The neural network model uses a residual module to process the oscillating data. The residual module generates output features through a bypass structure, and the output features are represented as follows:

[0018] Y = f main (X)+f bypass (X)

[0019] In the formula, X and Y represent the module's input and output characteristics, respectively, and f main (·) and f bypass (·) represent the mapping of the main road and the bypass, respectively;

[0020] The neural network model generates channel weight coefficients through a channel attention mechanism, and the output features are represented as follows:

[0021] Y = C(X) * X

[0022] In the formula, C(X) is the channel weight coefficient, which is generated by the interaction of the feature X through the global average pooling (GAP) operation and one-dimensional convolution of each channel;

[0023] The neural network model introduces nonlocal modules to achieve global feature interaction. After global interaction, feature Y... g The relationship between the input feature X and the input feature X can be expressed as:

[0024] Y g =θ(X) T φ(X)g(X)

[0025] In the formula, θ(·), φ(·), and g(·) represent three independent convolutional interaction operations, and the final model output feature Y is formed by another residual connection operation:

[0026] Y = W(Y) g )+X

[0027] In the formula, W(·) represents another convolutional interaction mapping;

[0028] Based on the output characteristics, it is determined whether the single station is an oscillation source.

[0029] As described above, the single-site location and removal method for oscillation source stations based on meta-learning, wherein the meta-learning process for generating the optimal initial parameters of the neural network model includes:

[0030] A forced oscillation sample set is constructed, which is generated by injecting harmonic current into the simulation system;

[0031] The forced oscillation sample set is divided into multiple source task sets, which are used for meta-learning training.

[0032] The neural network model is updated by alternating between internal and external updates. The internal update uses the training set of each source task set to generate the gradient of the test set, and the external update sums the gradients of the test set to update the initial parameters, thereby generating the model with the optimal initial parameters.

[0033] The above-described meta-learning-based method for locating and disconnecting a single oscillation source at a power station, wherein if the single power station is an oscillation source, triggering a circuit breaker to perform a disconnection operation, includes:

[0034] Based on the judgment results of the neural network model, the oscillation source label of the single station is determined;

[0035] If the oscillation source tag is an oscillation source, then the circuit breaker of the single station is triggered to perform a disconnection operation;

[0036] The oscillation suppression effect is evaluated based on the total distortion rate of the phase A current at the substation outlet, where the total distortion rate is expressed as I. THD :

[0037]

[0038] In the formula, I 50 The 50Hz spectral amplitude of the A-phase current at the outlet of the collecting station after oscillation stabilization is obtained through FFT analysis. f This represents the spectral amplitude at frequency f in the FFT analysis results.

[0039] The single-site location and removal method for oscillation source stations based on meta-learning, as described above, includes gradient updates of the neural network model through internal and external update loops, comprising:

[0040] Internal update: For each source task set, a small number of gradient updates are performed using the training set to generate the gradient of the test set; the gradient of the test set of each source task set is then fed into the external update process.

[0041] External updates are performed by updating the initial parameters of the neural network model based on the summation of gradients from each test set.

[0042] Repeat the internal and external update loops until the optimal initial parameter model is generated.

[0043] The single-site location and removal method for oscillation source stations based on meta-learning, as described above, includes performing maximum-minimum normalization processing on the electrical characteristics, which includes:

[0044] For each channel's sample data, perform max-min normalization to map all channel features to the 0-1 interval. The max-min normalization method is as follows:

[0045]

[0046] In the formula, X i and X i The original data and normalized data for different channels in the sample are shown separately.

[0047] Another objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned meta-learning-based single-site location and removal method for oscillation source stations.

[0048] Another objective of this invention is to provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the meta-learning-based single-site location and removal method for oscillation source sites.

[0049] Beneficial effects:

[0050] (I) This invention proposes a single-site location and cutoff concept for machine learning oscillation source localization, avoiding data communication and transmission between the site and the collection station, while extracting operational data from the initial stage of oscillation for analysis. This invention can effectively improve the timeliness and engineering practicality of single-site location and cutoff for machine learning oscillation source localization.

[0051] (II) This invention establishes a meta-learning process for oscillation source localization. By using simple forced oscillation data samples from two field stations for meta-learning, the optimal initialization parameters of the oscillation localization model are obtained, and the learned knowledge is transferred to the actual subsynchronous / supersynchronous oscillation domain. This invention can effectively solve the problem of insufficient accuracy in oscillation source localization caused by the scarcity of subsynchronous / supersynchronous oscillation data.

[0052] (III) The positioning algorithm proposed in this invention exhibits high positioning accuracy and robustness in positioning tasks of wind farms of different sizes. At the same time, the algorithm has universal applicability when facing the system structure of wind farms of different sizes and can effectively suppress oscillations. Attached Figure Description

[0053] Figure 1 A schematic diagram of the oscillation data acquisition method of the present invention.

[0054] Figure 2 A schematic diagram of the overall structure of the neural network of this invention.

[0055] Figure 3 A schematic diagram of the MAML meta-learning process of this invention.

[0056] Figure 4 A schematic diagram of a simulation example of forced oscillation generation in this invention.

[0057] Figure 5 The forced current waveform diagram of this invention.

[0058] Figure 6 The present invention provides a flowchart of a single-site location and removal method for oscillation source stations based on meta-learning.

[0059] Figure 7 The topology diagram of the four wind farms connected to the weak AC network in this embodiment of the invention.

[0060] Figure 8 Feature distribution diagram of the original model in this embodiment of the invention.

[0061] Figure 9 M in the embodiments of the present invention best Feature distribution map.

[0062] Figure 10 A timeliness comparison diagram between the single-site method and the communication method proposed in this embodiment of the invention.

[0063] Figure 11The waveforms with / without oscillation source localization and removal are compared with FFT in the embodiments of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0065] This invention provides a single-site location and removal method for oscillation source stations based on meta-learning. The proposed algorithm only collects oscillation operation data measured locally at the outlet of a single station. Specifically, the active power, reactive power, phase A voltage, and phase A current at the station outlet are used as input features X of the oscillation location model. in , denoted as X in ={P,Q,U a ,I a Traditional oscillation source location methods require the collection and comprehensive analysis of operational data from various sites. Compared to traditional methods, the proposed method avoids data transmission between sites, eliminates communication difficulties, and is more practical and timely in engineering applications.

[0066] Collect oscillation characteristic operation data X in Then, the data is fed into a machine learning algorithm. The algorithm extracts features from the data to determine whether the power station is an oscillation source. If the determination is "it is an oscillation source," the circuit breaker trips, disconnecting the power station and suppressing oscillation propagation. Conversely, if the determination is "it is not an oscillation source," the circuit breaker does not trip, and the power station continues to operate in grid-connected mode.

[0067] The oscillation mechanism of a power station grid-connected system is highly complex. This invention defines an "oscillation source station" as the station with the largest oscillation contribution. Specifically, this is determined through a hybrid dynamic simulation experiment involving station removal during training. In the experiment, if removing the c-th station yields the most significant oscillation suppression effect compared to removing other stations, then the c-th station is designated as the "oscillation source station." The oscillation suppression effect is specifically evaluated using the Total Harmonics Distortion (THD) of the outlet current at the collecting station after removal. The Total Harmonics Distortion (THD) is defined as I... THD As shown in equation (1):

[0068]

[0069] In the formula, I 50 The 50Hz spectral amplitude of the A-phase current at the outlet of the collecting station after oscillation stabilization is obtained through FFT analysis. fThis represents the spectral amplitude at frequency f in the FFT analysis results. In this invention, the frequency f ranges from 4-100Hz within the sub- / supersynchronous frequency band. The total current distortion rate I after station disconnection. THD The smaller the value, the more significant the oscillation suppression effect.

[0070] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0071] This embodiment provides a single-site location and removal method for oscillation source stations based on meta-learning, referencing... Figure 6 This includes the following steps:

[0072] Acquire oscillation operation data for a single station, the oscillation operation data including the electrical characteristics of the station outlet;

[0073] The oscillation operation data is used to extract features by a pre-trained neural network model to determine whether the single station is an oscillation source;

[0074] The optimal initial parameters of the neural network model are generated based on a meta-learning process, which is trained using forced oscillating samples.

[0075] If the single station is an oscillation source, the circuit breaker is triggered to perform a disconnection operation to suppress the oscillation propagation.

[0076] For example.

[0077] Specifically, oscillation operation data of a single station is acquired, and the oscillation operation data includes the electrical characteristics of the station outlet.

[0078] To further improve the timeliness of location and site removal, this embodiment collects operational data in the initial stage after the oscillation occurs (such as...). Figure 1 As shown in the figure, the oscillation source can be located during the oscillation development stage.

[0079] Furthermore, feature extraction is performed on the oscillation operation data using a pre-trained neural network model to determine whether the single station is an oscillation source. Specifically, this includes:

[0080] The neural network model for oscillation localization proposed in this invention is built on the traditional convolutional-fully connected layer model. It uses residual modules to replace convolutional networks and introduces two types of network modules: channel attention and nonlocal modules, in order to improve localization accuracy.

[0081] Residual Module: In the initial stage of oscillation, the connection between the running data and the localization information may be weak, making it difficult for simple shallow networks to fully extract features. When using deep networks to mine features, the gradient vanishing problem may occur. The residual module introduces a bypass structure to avoid this phenomenon. This invention uses a residual module to replace the traditional convolutional module. The relationship between the module's input and output can be expressed as:

[0082] Y = f main (X)+f bypass (X) (2)

[0083] In the formula, X and Y represent the module's input and output characteristics, respectively, and f main (·) and f bypass (·) represent the mapping of the main road and the bypass road, respectively.

[0084] Channel Attention Mechanism: From a physical perspective, different electrical features are correlated with oscillation localization problems to varying degrees. Guiding the model to focus more on important information in the channel dimension during feature extraction can further improve model performance. Therefore, this invention introduces an Efficient Channel Attention (ECA) module. The relationship between the output feature Y and the input feature X of the ECA module can be expressed as:

[0085] Y = C(X) * X (3)

[0086] In the formula, C(X) is the channel weight coefficient, which is generated by the interaction of global average pooling (GAP) and one-dimensional convolution of each channel for feature X.

[0087] Non-local modules: In oscillatory problems, the running data exhibits periodic changes. Traditional convolution operations can only perform local interactions between adjacent elements; however, achieving global interaction of oscillating data elements can further explore the periodic features of the oscillating data. This invention introduces non-local modules to achieve global feature interaction. After global interaction, feature Y... g The relationship between the input feature X and the input feature X can be expressed as:

[0088] Y g =θ(X) T φ(X)g(X) (4)

[0089] In the formula, θ(·), φ(·), and g(·) represent three independent convolutional interaction operations. The final model output feature Y is formed by another residual connection operation:

[0090] Y = W(Y) g )+X (5)

[0091] In the formula, W(·) represents another convolutional interaction mapping.

[0092] This invention incorporates the above three modules into the feature extraction process, and designs a neural network model structure for oscillating single-station positioning as follows: Figure 2 As shown.

[0093] Furthermore, the optimal initial parameters of the neural network model are generated based on a meta-learning process, which utilizes forced oscillatory samples for training. Specifically, this includes:

[0094] The scarcity of real-world subsynchronous / supersynchronous oscillation data is a significant challenge for machine learning algorithms applied to the field of oscillation. Training machine models based solely on a limited number of real-world oscillation samples easily leads to overfitting and reduced model localization accuracy. This invention proposes a model-agnostic meta-learning (MAML) method for single-site localization of oscillation sources. This method uses forced oscillation samples to construct the meta-learning process. Although forced oscillations and subsynchronous / supersynchronous oscillations have different initiation mechanisms, they both fall under the category of oscillation problems and share similarities in data characteristics, exhibiting multi-frequency coupled periodic variations. Furthermore, forced oscillation scenarios are easier to obtain than real-world subsynchronous / supersynchronous oscillation scenarios. Therefore, this invention selects the forced oscillation sample domain as the source task for meta-learning to address the problem of scarce oscillation data.

[0095] The core idea of ​​meta-learning for oscillation source localization is to teach the machine model how to efficiently mine oscillation source-related information. When faced with a new subsynchronous / hypersynchronous oscillation source localization task, the model can quickly converge to a high accuracy with only a small number of gradient updates. The goal of MAML is to enable the model to possess a set of optimal initial parameters θ through a meta-learning process involving multiple sets of forced oscillation source localization tasks (referred to as source tasks in this invention). best When facing the localization of subsynchronous / supersynchronous oscillation sources (referred to as the target task in this invention), the model starts from θ. best Gradient updates are performed based on this foundation, relying on only a small number of new task samples to achieve high localization performance. The MAML algorithm does not depend on any specific neural network model structure and has good universality.

[0096] A schematic diagram of the MAML algorithm's meta-learning process is shown below. Figure 3 As shown.

[0097] Unlike traditional deep learning gradient update methods, the MAML source task meta-learning process introduces two gradient update loops: "internal update" and "external update." In the "internal update," for each specific source task i (i∈[1,N)... s ], N s (where θ is the number of source tasks), the model performs a small number of gradient updates using the task training set, and updates the gradients ▽θ' of the test set.i The data is fed into the "external update" process. In the "external update" process, the gradients ▽θ'i of each task's test set are summed, and the summed gradient Σθ' is used to update the model's initial parameters. Through continuous inner and outer layer update processes, the optimal initial parameters θ are finally obtained. best The corresponding network model is called the optimal initial parameter model M. best .

[0098] This invention constructs a simple simulation example of two direct-drive wind farms connected to a weak AC grid to generate forced oscillation samples, such as... Figure 4 As shown. By injecting small-signal harmonic current disturbances at the outlet busbars of the two stations respectively, forced oscillation samples with oscillation sources of station 1 and station 2 can be obtained. The source task of this invention only performs single-station positioning for station 1, constructing a station 1 positioning task sample set S = {(x i ,y i )|i∈n s}, where xi and yi are the feature data of station 1 and the corresponding label indicating whether station 1 is an oscillation source, respectively, and n s The sample set is randomly divided into N groups, representing the total number of samples. s Groups are formed to create multiple source task sets required for meta-learning, denoted as S1, S2, ... S. Ns After acquiring multiple source task sets, MAML meta-learning is performed to obtain the optimal initial parameter model M for oscillatory localization. best .

[0099] Compared to network models initialized with random parameters, the optimal initial parameter model M is obtained through forced oscillation training. best It already possesses a certain ability to mine oscillation data features and identify oscillation sources. Based on this, training for sub / supersynchronous oscillation localization tasks can achieve high accuracy with only a small number of samples and iterations.

[0100] This invention induces forced oscillations by injecting harmonic current at the wind farm outlet in a simulation system. To further enhance the similarity between the forced oscillations and actual oscillations, this invention considers the frequency symmetry coupling of actual subsynchronous / supersynchronous oscillations and the time-varying amplitude of the initial stage of oscillation in the harmonic current, ultimately injecting the harmonic current as follows: Figure 5 As shown.

[0101] Furthermore, if the single substation is an oscillation source, the circuit breaker is triggered to perform a disconnection operation to suppress oscillation propagation. The specific method for locating and disconnecting a single substation as an oscillation source includes:

[0102] The specific application method of the subsynchronous / supersynchronous oscillation single-site positioning and removal algorithm proposed in this invention can be divided into two parts: "offline training" and "online application".

[0103] In the "Offline Training" step:

[0104] First, a two-station system is constructed, and a forced oscillation sample set for station 1 is obtained. This sample set is then used to perform a meta-learning process on the neural network model to obtain the optimal initial parameter model M. best In the meta-learning process, the number of source task sets N is divided. s =9, using the Adam optimizer for gradient updates, the model loss function is cross-entropy, the number of internal update iterations is 1, and the number of external update iterations is 100.

[0105] (2) Construct a simulation model of the actual system and obtain the subsynchronous / supersynchronous oscillation dataset. Through data processing and cutoff experiments, obtain the sample set T for station c. c ={(x j,c ,y j,c )},x j,c With y j,c Do not collect data at the exit of station c for the j-th sample, and determine whether station c is an oscillation source label.

[0106] (3) The optimal initial parameter model M best Duplicate N times (N is the total number of stations), and use their respective sample sets T. c For the optimal initial parameter model M best A second gradient update is performed, ultimately yielding the single-site positioning model M for each site. c The gradient update process still uses the Adam optimizer, and the model loss function is cross-entropy.

[0107] In the "Online Application" step:

[0108] (1) The single-site positioning model M obtained from "offline training" for each site c The data is translated and stored in the corresponding control and protection program of the cth station in the actual project.

[0109] (2) When the system oscillates, the online measurement devices at each station collect oscillation data and send it to their respective positioning models M. c The program analyzes the data to determine if the c-th power station is an oscillation source. If the c-th power station is determined to be an oscillation source, a cutoff operation is performed on the c-th power station to quickly suppress the oscillation.

[0110] To further verify the practical effects of the present invention, specific application embodiments are also provided.

[0111] Example 1

[0112] Build such a simulation system in PSCAD / EMTDC Figure 7The four power station grid connection examples shown demonstrate how this method constructs 324 subsynchronous / supersynchronous oscillation samples by continuously adjusting the output of each station and setting the station parameters. All samples originate from natural resonance, and the oscillations are caused by energy exchange between the control loop or system capacitors / inductors. The subsynchronous / supersynchronous oscillation frequency ranges from 5-12Hz to 88-95Hz, and the total distortion rate of the A-phase current at the collector station outlet ranges from 0.3018 to 0.6153.

[0113] For each sample, oscillation source labels were set using a site removal experiment, referencing the oscillation source definition method. This ultimately generated a sample set T for each site. c ={(x j,c ,y j,c The experimental sampling window length of this method is 0.2s, the sampling frequency is 2kHz, and the time sequence length of the electrical characteristic variables is 400 (0.2s×2kHz). To achieve more periodic information interaction, this method transforms the electrical runtime sequence data into a [10,40] two-dimensional matrix form.

[0114] Sample set T c The dataset was randomly divided into training and test sets in a 4:1 ratio. To test the accuracy and robustness of the proposed algorithm when samples are scarce, the training set was further randomly divided into 9 parts for training the algorithm separately. The accuracy was taken as the average of the corresponding test set values. A total of 5177 forced oscillation samples were generated, with the forced current frequency ranged from 4 to 99 Hz. These forced oscillation data samples were then divided into 9 parts, representing the number of source tasks N. s =9.

[0115] To eliminate the influence of data dimensions between different samples and channels and accelerate the subsequent gradient descent of the neural network, max-min normalization is performed during the data preprocessing stage.

[0116] The performance accuracy of the positioning model is determined by the accuracy A. cc Given:

[0117]

[0118] Where T is the total number of samples, TP is the number of samples where both the model prediction and the oscillation source label are "oscillation source", and TN is the number of samples where both the model prediction and the oscillation source label are "non-oscillation source".

[0119] The positioning accuracy (Acc) results of the proposed algorithm are shown in Table 1. As can be seen from Table 1, the algorithm achieves an accuracy (Acc) of [missing information - likely related to positioning accuracy at four different sites]. cc The average accuracy reaches 99.02%, demonstrating high positioning accuracy that meets engineering requirements. Furthermore, the algorithm achieves a positioning accuracy of over 97% for each site, proving its robustness.

[0120] Table 1 shows the accuracy of the proposed algorithm for location determination at various stations.

[0121]

[0122] To verify the effectiveness of the proposed forced oscillation meta-learning process on positioning accuracy, an algorithm without meta-learning training was set up as a control experiment. The relevant experimental results are shown in Table 2. Comparing the results in Tables 1 and 2, it can be seen that after introducing the meta-learning process, the positioning performance of the algorithm at each station is significantly improved. In addition, in the algorithm without the meta-learning process, the positioning accuracy of each station is very unstable, with an accuracy gap of 27.35%, which is difficult to meet the needs of engineering applications; however, after introducing the meta-learning process, the positioning accuracy gap between different stations is significantly reduced to 1.21%, proving that the meta-learning-based positioning algorithm has good robustness and is more suitable for engineering applications.

[0123] Table 2 shows the accuracy of location judgment at each station during the meta-learning process.

[0124]

[0125] To further demonstrate the effectiveness of the meta-learning process, the t-SNE algorithm was used to visualize and analyze the feature distribution before the last classifier layer. Figure 8 This shows the sample feature distribution corresponding to the original model with randomly initialized parameters. Figure 9 This demonstrates the optimal initial parameter model M after meta-learning. best The characteristic distribution is shown, where samples from different oscillation sources are represented by different colors.

[0126] by Figure 8 (a) is an example. Figure 8 (a) shows the feature distribution of data from different oscillation source samples at station 1 after feature extraction using the original model. For the station 1 localization task, samples with station 1 as the oscillation source (red samples in the figure) correspond to the "oscillation source" samples described in the single-station localization method, while the other three types of samples (green, blue, and purple samples in the figure) correspond to "non-oscillation source" samples. (Comparison) Figure 8 (a) and Figure 9 (a) It can be found that, Figure 8 In (a), the red sample has a large overlap with the distribution of the other three types of samples, while Figure 9 In (a), the red sample is significantly more distant from the other three types of samples. Figure 8 Based on (a), a large number of sample updates are still needed to distinguish between "oscillation source" and "non-oscillation source" samples. Figure 9 (a) It already has a good foundation in classification and only requires a small amount of training to perform accurate classification.

[0127] Comparison Figure 8 (b)-(d) and Figure 9 (b)-(d) yield the same conclusion as above. Therefore, it can be proved that M best Each site can, to some extent, distinguish between samples that are "oscillation sources" and those that are "non-oscillation sources," laying a good foundation for subsequent accurate classification.

[0128] To verify the effectiveness of this method, which incorporates a residual module, channel attention mechanism, and nonlocal module into the model structure, an ablation experiment was conducted. The average and minimum positioning accuracies of four stations were compared to demonstrate the impact of the three modules on positioning accuracy. The ablation results are shown in Table 3. The results in the table show that, regardless of the average or minimum accuracy, the added residual module, channel attention module, and nonlocal module all positively enhance the model's accuracy. The fusion of these three modules can further improve the model's positioning accuracy and robustness.

[0129] Table 3 Ablation Experiment Results

[0130]

[0131] To verify the performance of the proposed algorithm for localization, a comparative experiment was conducted, comparing the proposed algorithm with existing machine learning oscillatory localization methods, including Graph Convolutional Network (GCN), autoencoder + Long Short-Term Memory (LSTM), K-Nearest Neighbors (KNN), and model fine-tuning transfer learning algorithms. The model comparison results are shown in Table 4. The results demonstrate that the proposed single-site localization method outperforms all the aforementioned existing machine learning methods.

[0132] Table 4 Comparison of Results from Different Algorithms

[0133]

[0134] To demonstrate the timeliness of the proposed algorithm, the time required for the algorithm during "online application" was tested. The calculation program was built using Python based on the PyTorch framework, and the computer processor was an Intel(R) Core(TM) i7-8700 CPU@3.20GHz.

[0135] After testing, the proposed single-site positioning and removal algorithm can complete the oscillation source removal in only about 404ms after the oscillation source occurs. This includes 200ms of oscillation operation required for the sampling window length, 37ms required for data reading and positioning model calculation, 100ms for circuit breaker operation, and 20% margin time.

[0136] In addition, the application time of traditional site interactive positioning and cut-off methods that require communication is calculated, such as... Figure 10 As shown in the figure, during the communication process, assuming the maximum transmission frequency from the field station to the aggregation station is 10ms / frame, and the amount of transmitted data is compressed to one-tenth of the original sampled data, the required communication time is 400ms. In comparison, the proposed single-field station positioning time is reduced by 54.20%, demonstrating better timeliness.

[0137] To demonstrate the effectiveness of the proposed single-site positioning and removal algorithm in suppressing oscillations, Figure 11 This paper presents a comparison of the A-phase current waveform at the substation outlet with FFT results for a sample under two conditions: without removal of the oscillation source and with removal of the oscillation source. The two conditions are represented by yellow and blue, respectively. Oscillation occurs at 1.2s, and the oscillation source is removed after 404ms. FFT analysis is performed on the data after the oscillation stabilizes. Both the waveform and the FFT spectrum clearly demonstrate that the proposed oscillation source localization and removal method can effectively suppress oscillation.

[0138] Furthermore, the total distortion rate of the A-phase current at the collection station outlet was calculated for all 324 samples. Without targeted cutoff, the average total distortion rate was 0.4768. After targeted cutoff using the proposed algorithm, the average total distortion rate became 0.0164, a reduction of 96.56%, demonstrating a significant oscillation suppression effect.

[0139] Example 2:

[0140] To verify the applicability of the proposed algorithm to systems of different station sizes, six station grid-connected simulation examples were constructed in the PSCAD / EMTDC simulation system, obtaining a total of 96 samples. The model details are the same as in Implementation Example 1. The positioning accuracy A of the proposed algorithm in the six station systems is... cc The results are shown in Table 5.

[0141] Table 5. Positioning accuracy of six station systems

[0142]

[0143] As shown in Table 5, the algorithm achieves an accuracy of A in single-site positioning tasks at six sites. cc The average accuracy reached 97.50%, and the minimum accuracy exceeded 94%, indicating that the positioning accuracy still meets engineering requirements. Without positioning cutoff, the total distortion rate of the outlet current at the collection station was 0.3564, while after positioning cutoff, the distortion rate was 0.0053, a reduction of 98.51%. Therefore, it can be proven that the proposed single-site positioning method has high applicability for locating oscillation sources in wind power grid connections of different scales.

[0144] Example 3

[0145] This embodiment of an oscillating single-site positioning and cutting electronic device and storage device based on meta-learning includes:

[0146] One or more processors on which one or more programs run;

[0147] A storage device on which one or more programs are stored;

[0148] When the one or more programs are executed by the one or more processors, the one or more processors implement the meta-learning-based single-site location and removal method for oscillation source stations; the memory is used to store the computer program for the meta-learning-based single-site location and removal method for oscillation source stations.

[0149] Alternatively, the aforementioned electronic device may be a server.

[0150] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A single-site location and removal method for oscillation source stations based on meta-learning, characterized in that, Includes the following steps: Obtain oscillation operation data from a single station, preprocess the oscillation operation data, and obtain input features; The oscillation operation data is used to extract features by a pre-trained neural network model to determine whether the single station is an oscillation source; The optimal initial parameters of the neural network model are generated based on a meta-learning process, which is trained using forced oscillating samples. If the single station is an oscillation source, the circuit breaker is triggered to perform a disconnection operation to suppress the oscillation propagation.

2. The method as described in claim 1, characterized in that, The acquisition of oscillation operation data for a single station includes: The measuring device of the single station collects the electrical characteristics of the station outlet, including active power, reactive power, phase A voltage and phase A current; The electrical features are subjected to max-min normalization to generate normalized input features.

3. The method as described in claim 1, characterized in that, The step of extracting features from the oscillation operation data using a pre-trained neural network model to determine whether a single station is an oscillation source includes: The neural network model uses a residual module to process the oscillating data. The residual module generates output features through a bypass structure, and the output features are represented as follows: Y=f main (X)+f bypass (X) In the formula, X and Y represent the module's input and output characteristics, respectively, and f main (·) and f bypass (·) represent the mapping of the main road and the bypass, respectively; The neural network model generates channel weight coefficients through a channel attention mechanism, and the output features are represented as follows: Y = C(X) * X In the formula, C(X) is the channel weight coefficient, which is generated by the interaction of global average pooling and one-dimensional convolution of feature X through each channel. The neural network model introduces nonlocal modules to achieve global feature interaction. After global interaction, feature Y... g The relationship between the input feature X and the input feature X can be expressed as: Y g =θ(X) T φ(X)g(X) In the formula, θ(·), φ(·), and g(·) represent three independent convolutional interaction operations, and the final model output feature Y is formed by another residual connection operation: Y=W(Y g )+X In the formula, W(·) represents another convolutional interaction mapping; Based on the output characteristics, it is determined whether the single station is an oscillation source.

4. The method as described in claim 1, characterized in that, The process of generating the optimal initial parameters of the neural network model based on meta-learning includes: A forced oscillation sample set is constructed, which is generated by injecting harmonic current into the simulation system; The forced oscillation sample set is divided into multiple source task sets, which are used for meta-learning training. The neural network model is updated by alternating between internal and external updates. The internal update uses the training set of each source task set to generate the gradient of the test set, and the external update sums the gradients of the test set to update the initial parameters, thereby generating the model with the optimal initial parameters.

5. The method as described in claim 1, characterized in that, If the single station is an oscillation source, the circuit breaker is triggered to perform a disconnection operation, including: Based on the judgment results of the neural network model, the oscillation source label of the single station is determined; If the oscillation source tag is an oscillation source, then the circuit breaker of the single station is triggered to perform a disconnection operation; After the switch-out operation is completed, the oscillation suppression effect is evaluated based on the total distortion rate of the A-phase current at the collection station outlet, where the total distortion rate is expressed as I. THD : In the formula, I 50 The 50Hz spectral amplitude of the A-phase current at the outlet of the collecting station after oscillation stabilization is obtained through FFT analysis. f This represents the spectral amplitude at frequency f in the FFT analysis results.

6. The method as described in claim 3, characterized in that, The gradient update of the neural network model through internal and external update loops includes: Internal update: For each source task set, a small number of gradient updates are performed using the training set to generate the gradient of the test set; the gradient of the test set of each source task set is then fed into the external update process. External updates are performed by updating the initial parameters of the neural network model based on the summation of gradients from each test set. Repeat the internal and external update loops until the optimal initial parameter model is generated.

7. The method as described in claim 2, characterized in that, The step of performing maximum-minimum normalization on the electrical characteristics includes: For each channel's sample data, perform max-min normalization to map all channel features to the 0-1 interval. The max-min normalization method is as follows: In the formula, X i and X i The original data and normalized data for different channels in the sample are shown separately.

8. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the meta-learning-based single-site location and removal method for oscillation source sites according to any one of claims 1-7.

9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program and the processor runs the computer program to cause the electronic device to perform the meta-learning-based single-site location and removal method for oscillation source sites as described in any one of claims 1-7.

Images