Wind power plant current oscillation control method

Abstract

The invention discloses a wind power plant current oscillation control method, and belongs to the technical field of power system automation. Comprising the following steps: constructing a data set and an oscillation prediction network, and training and testing the oscillation prediction network according to the data set; obtaining a target operation parameter sequence and a target space topological graph of the target wind power plant, and processing the target operation parameter sequence and the target space topological graph according to the oscillation prediction network to obtain a prediction result; and the control system of the target wind power plant performs parameter adjustment on equipment of the target wind power plant according to the prediction result. According to the method, the target operation parameters and sequence of the target wind power plant and the target space topological graph are obtained, oscillation prediction of the target wind power plant is achieved, then the equipment parameters of the target wind power plant are adjusted in time according to the prediction result, and timely and accurate prediction and suppression of current oscillation of the target wind power plant are achieved.

Description

A method for controlling current oscillation in a wind farm Technical Field

[0001] This application belongs to the field of power system automation technology, specifically relating to a method for controlling current oscillations in wind farms. Background Technology

[0002] Current oscillations in wind farms mainly originate from factors such as random wind speed fluctuations, inter-generator coupling interference, grid impedance changes, and high-frequency switching of power electronic equipment. These oscillations manifest as periodic or non-periodic fluctuations in current amplitude and frequency. Such oscillations not only lead to unstable wind turbine output and increased equipment losses, but can also cause grid voltage fluctuations, relay protection malfunctions, and in severe cases, even large-scale power outages, directly threatening the safe and stable operation of the power system.

[0003] Traditional control methods, such as PID control, adaptive damping control, and model predictive control (MPC), are based on mathematical models of wind farms and power systems to design control strategies. However, traditional control methods have a lag in response and are difficult to cope with rapidly changing oscillating conditions. Summary of the Invention

[0004] The purpose of this application is to provide a wind farm current oscillation control method that can solve the technical problem that the control response in the prior art is lagging and it is difficult to cope with rapidly changing oscillation conditions.

[0005] To address the aforementioned technical problems, this application provides the following: Firstly, embodiments of this application offer a method for controlling current oscillations in a wind farm. The method includes: acquiring operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset based on these sequences; constructing an oscillation prediction network, which includes a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; training the oscillation prediction network using a portion of the dataset, and testing the trained network using another portion of the dataset; acquiring a target operating parameter sequence and a target spatial topology map of a target wind farm, processing the target operating parameter sequence and the target spatial topology map using the oscillation prediction network to obtain a prediction result; and adjusting the parameters of the equipment in the target wind farm based on the prediction result using the control system of the target wind farm.

[0006] As an optional implementation of the first aspect of this application, the step of training the oscillation prediction network based on a portion of the dataset and testing the trained oscillation prediction network based on another portion of the dataset specifically involves: dividing the dataset into a training set and a test set according to a preset ratio, and constructing a mean squared error loss function; iteratively training the oscillation prediction network based on the training set and the mean squared error loss function until the parameters of the oscillation prediction network converge; testing the oscillation prediction network after parameter convergence based on the test set to obtain test results, and evaluating the test results based on a generalization ability evaluation index to obtain evaluation results; if the evaluation results are satisfactory, the training and testing of the oscillation prediction network are completed; if the evaluation results are unsatisfactory, the parameters of the oscillation prediction network are adjusted and retrained until the evaluation results are satisfactory.

[0007] As an optional implementation of the first aspect of this application, the oscillation prediction network processes the target operating parameter sequence and the target spatial topology map to obtain a prediction result. Specifically, the network processes the target operating parameter sequence according to the temporal feature extraction and enhancement module to obtain temporal attention features; processes the target spatial topology map according to the spatial feature extraction and enhancement module to obtain spatial attention features; fuses the temporal attention features and the spatial attention features according to the spatiotemporal feature fusion module to obtain spatiotemporal attention features; and performs prediction processing on the spatiotemporal attention features according to the multi-task prediction head to obtain the prediction result. The prediction result includes the predicted oscillation amplitude and predicted oscillation frequency of the target wind farm in a future first time period. As an optional implementation of the first aspect of this application, the temporal feature extraction and enhancement module performs temporal feature extraction processing on the target operating parameter sequence to obtain temporal attention features. Specifically, it performs long-term temporal feature extraction on the target operating parameter sequence according to the long-term feature extraction unit to obtain target long-term temporal features; performs short-term temporal feature extraction on the target operating parameter sequence according to the short-term feature extraction unit to obtain target short-term temporal features; concatenates the target long-term temporal features and the target short-term temporal features along the channel dimension to obtain long-short-term fusion features; and performs attention enhancement processing on the long-short-term fusion features according to the temporal attention mechanism to obtain the temporal attention features. As an optional implementation of the first aspect of this application, the step of extracting long-term features from the target operating parameter sequence using the long-term feature extraction unit to obtain target long-term features specifically involves: processing the target operating parameter sequence using a bidirectional long short-term memory network layer to obtain a first long-term memory feature; performing self-attention enhancement processing on the first long-term memory feature using a self-attention mechanism to obtain a long-term attention-enhanced feature; processing the long-term attention-enhanced feature using a bidirectional long short-term memory network layer to obtain a second long-term memory feature; and performing global average pooling processing on the second long-term memory feature using a pooling layer to obtain the target long-term features.As an optional implementation of the first aspect of this application, the step of extracting short-time features from the target running parameter sequence using a short-time feature extraction unit to obtain target short-time features specifically involves: performing convolution processing on the target running parameter sequence using a 1×1 convolutional layer to obtain a first short-time convolutional feature; performing convolution processing on the first short-time convolutional feature using 1×3, 1×5, and 1×7 convolutional layers respectively to obtain a second, third, and fourth short-time convolutional feature respectively; and then convolving the second, third, and fourth short-time convolutional features along... The channel dimension is concatenated to obtain the first short-temporal fusion feature; the first short-temporal fusion feature is then enhanced with attention according to the self-attention mechanism to obtain the short-temporal attention-enhanced feature; the short-temporal attention-enhanced feature is then convolved sequentially by multiple 1×3 convolutional layers to obtain the fifth short-temporal convolutional feature output by each of the multiple 1×3 convolutional layers; the fifth short-temporal convolutional features output by each of the 1×3 convolutional layers are then element-wise summed to obtain the second short-temporal fusion feature; the first short-temporal fusion feature and the second short-temporal fusion feature are then concatenated along the channel dimension to obtain the target short-temporal feature. As an optional implementation of the first aspect of this application, the step of performing attention enhancement processing on the long-short temporal fusion features according to the temporal attention mechanism to obtain the temporal attention features specifically involves: performing global average pooling processing on the long-short temporal fusion features according to a pooling layer to obtain long-short temporal average pooling features; performing full connection processing on the long-short temporal average pooling features according to a fully connected layer to obtain long-short temporal fully connected features; performing Softmax activation processing on the long-short temporal fully connected features according to an activation layer to obtain temporal attention weights; and performing element-wise multiplication of the temporal attention weights with the long-short temporal fusion features to obtain the temporal attention features.

[0008] Secondly, embodiments of this application provide a wind farm current oscillation control system, the system comprising: a data construction module: acquiring operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset based on the operating parameter sequences and spatial topology maps of multiple wind farms; a network construction module: constructing an oscillation prediction network, the oscillation prediction network including a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; a network training module: training the oscillation prediction network based on a portion of the dataset, and testing the trained oscillation prediction network based on another portion of the dataset; a first processing module: acquiring a target operating parameter sequence and a target spatial topology map of a target wind farm, processing the target operating parameter sequence and the target spatial topology map based on the oscillation prediction network to obtain a prediction result; and a second processing module: the control system of the target wind farm adjusts the parameters of the equipment in the target wind farm based on the prediction result.

[0009] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0010] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0011] In this embodiment, compared with the prior art, the following technical effects are achieved: This application obtains the target operating parameters and sequence and the target spatial topology map of the target wind farm to predict the oscillation of the target wind farm, and then adjusts the equipment parameters of the target wind farm in a timely manner according to the prediction results, thereby achieving timely and accurate prediction and suppression of the current oscillation of the target wind farm; the oscillation prediction network achieves dual extraction of temporal and spatial attention features through a temporal and spatial dual attention mechanism composed of a temporal feature extraction enhancement module and a spatial feature extraction enhancement module; the spatiotemporal feature fusion module fuses the temporal and spatial attention features through a cross-attention mechanism to obtain the final spatiotemporal attention features; finally, the spatiotemporal attention features are processed by a multi-task prediction head to achieve high-precision prediction of the oscillation prediction network; the temporal feature extraction enhancement module extracts the target long-time-series features and the target short-time-series features through feature extraction by a long-time-series feature extraction unit and a short-time-series extraction unit; the target long-time-series features and the target short-time-series features are fused through a splicing operation; finally, the fused long-short-time-series features are processed by a temporal attention mechanism to obtain... The temporal attention feature significantly enhances the feature capture capability of the temporal feature extraction enhancement module, thereby improving the reliability of network prediction. The long temporal feature extraction unit inserts a self-attention mechanism between two bidirectional long short-term memory networks, forming an information extraction enhancement process of temporal extraction, attention enhancement, and then temporal extraction, thereby extracting a highly continuous second long temporal memory feature. Global average pooling is then applied to the second long temporal memory feature to obtain the globally statistical target long temporal feature. This significantly improves the network's ability to extract long temporal information, thereby improving the network's final prediction accuracy. The short temporal feature extraction unit obtains a first short-term fusion feature through multi-scale information capture and fusion. This first short-term fusion feature is then subjected to multi-convolutional depth extraction, and features extracted at different depths are added and fused to obtain a second short-term fusion feature rich in information of different depths. Finally, the first short-term fusion feature from the multi-scale path and the second temporal fusion feature from the depth path are fused to achieve the fusion of multi-scale and depth information. This process significantly improves the network's ability to capture short temporal information, thereby improving the reliability and accuracy of the network's final prediction results. Attached Figure Description

[0012] Figure 1 is a flowchart of a wind farm current oscillation control method provided in some embodiments of this application; Figure 2 is a schematic diagram of the structure of the oscillation prediction network in a wind farm current oscillation control method provided in some embodiments of this application; Figure 3 is a schematic diagram of the structure of the time sequence feature extraction and enhancement module in a wind farm current oscillation control method provided in some embodiments of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0015] The following description, in conjunction with the accompanying drawings, details a wind farm current oscillation control method provided in this application through specific embodiments and application scenarios.

[0016] An embodiment of a wind farm current oscillation control method includes the following steps: S100: Acquire the operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset based on the operating parameter sequences and spatial topology maps of multiple wind farms; it should be noted that, through the monitoring and data acquisition system of each wind farm, the operating parameter sequences of each wind farm are acquired in real time, including the current, voltage, active power, reactive power, frequency, generator speed, pitch angle, etc. of each wind turbine and collection point, and time series data are formed at fixed time intervals. Each time series data is subjected to noise reduction, missing value completion, and normalization processing to obtain the operating parameter sequence corresponding to each wind farm. The electrical connection relationship and distance between the wind farm cluster power generation equipment are obtained through the power system dispatch center. A spatial topology map with graph structure is constructed with the cluster power generation equipment as nodes and the electrical connection relationship as edges; a set of data corresponding to the spatial topology map and operating parameter sequence of the same wind farm; a dataset is constructed based on the set of data corresponding to each wind farm.

[0017] S200: Construct an oscillation prediction network, which includes a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; S300: Train the oscillation prediction network based on a portion of the dataset, and test the trained oscillation prediction network based on another portion of the dataset; As an optional implementation of this embodiment, S300 specifically includes: S310: Divide the dataset into a training set and a test set according to a preset ratio, and construct a mean squared error loss function; S320: Iteratively train the oscillation prediction network based on the training set and the mean squared error loss function until the parameters of the oscillation prediction network converge; S330: Test the oscillation prediction network after parameter convergence based on the test set, obtain test results, and evaluate the test results based on the generalization ability evaluation index to obtain evaluation results; S340: If the evaluation results are qualified, the training and testing of the oscillation prediction network are completed; if the evaluation results are unqualified, adjust the parameters of the oscillation prediction network and retrain until the evaluation results are qualified.

[0018] It should be noted that the dataset is divided into training and test sets in a 4:1 ratio. A subset of data from the training set is used as a validation set. A mean squared error loss function is constructed. The network is trained using the selected training set and the mean squared error loss function. Training epochs are set, and the difference between the predicted value and the true label (i.e., the loss value) is calculated using the mean squared error loss function. The Adam optimizer is used to calculate the gradient of the loss value with respect to the weights of each layer in the network, and the weights are continuously updated to reduce the loss. This training is repeated for multiple epochs. When the loss on both the training and validation sets no longer decreases, the network is considered "test set". On the test set, the network's performance reaches stability, indicating that the parameters of the oscillation prediction network have converged. Next, the generalization ability of the oscillation prediction network is evaluated using the test set. The oscillation prediction network with converged input parameters is tested to obtain test results, which are then evaluated according to generalization ability evaluation metrics, including accuracy, precision, recall, and F1 score. If the evaluation results are satisfactory, the training and testing of the oscillation prediction network are completed. If the evaluation results are unsatisfactory, the parameters of the oscillation prediction network are adjusted and retrained until the evaluation results are satisfactory.

[0019] S400: Obtain the target operating parameter sequence and target spatial topology map of the target wind farm. Process the target operating parameter sequence and target spatial topology map according to the oscillation prediction network to obtain the prediction result. As an optional implementation of this embodiment, the oscillation prediction network in S400 processes the target operating parameter sequence and target spatial topology map to obtain the prediction result, specifically as follows: S410: Process the target operating parameter sequence according to the temporal feature extraction and enhancement module to obtain temporal attention features; S420: Process the target spatial topology map according to the spatial feature extraction and enhancement module to obtain spatial attention features; S430: Fuse the temporal attention features and spatial attention features according to the spatiotemporal feature fusion module to obtain spatiotemporal attention features; S440: Perform prediction processing on the spatiotemporal attention features according to the multi-task prediction head to obtain the prediction result; wherein, the prediction result includes the predicted oscillation amplitude and predicted oscillation frequency of the target wind farm in the first time period in the future.

[0020] It should be noted that, firstly, the temporal feature extraction and enhancement module processes the target operating parameter sequence to obtain temporal attention features; simultaneously, the spatial feature extraction and enhancement module processes the target spatial topology graph to obtain spatial attention features. The spatial feature extraction module uses a Graph Attention Network (GAT), which assigns different influence weights to the neighbors of each node in the target spatial topology graph. This allows us to obtain the dynamic, non-homogeneous mutual influence strength between nodes in the target spatial topology graph under a specific oscillation mode, thus obtaining spatial attention features. Then, the spatiotemporal feature fusion module fuses the temporal and spatial attention features to obtain spatiotemporal attention features. This spatiotemporal feature fusion module uses a cross-attention mechanism, allowing the temporal attention features to query which spatial relationships in the spatial attention features are more important in a specific time pattern, thus obtaining spatiotemporal attention features. Finally, the multi-task prediction head... The spatiotemporal attention features are processed to obtain the prediction results, namely the predicted oscillation amplitude and frequency in the first time interval of the future. The multi-task prediction head consists of multiple parallel connection layers. In this embodiment, there are two prediction heads, which are used to predict the oscillation amplitude and frequency respectively. The prediction heads take the spatiotemporal attention features as input and directly regress the specific physical quantities, namely the predicted oscillation amplitude and frequency in the first time interval of the future. The oscillation prediction network realizes the dual extraction of temporal and spatial attention features through a temporal and spatial dual attention mechanism composed of a temporal feature extraction enhancement module and a spatial feature extraction enhancement module. The spatiotemporal feature fusion module fuses the temporal and spatial attention features through a cross-attention mechanism to obtain the final spatiotemporal attention features. Finally, the spatiotemporal attention features are processed by the multi-task prediction head to achieve high-precision prediction of the oscillation prediction network.

[0021] As an optional implementation of this embodiment, the temporal feature extraction and enhancement module in S410 performs temporal feature extraction processing on the target running parameter sequence to obtain temporal attention features. Specifically: S411: The target running parameter sequence is subjected to long temporal feature extraction by the long temporal feature extraction unit to obtain target long temporal features; S412: The target running parameter sequence is subjected to short temporal feature extraction by the short temporal feature extraction unit to obtain target short temporal features; S413: The target long temporal features and target short temporal features are concatenated along the channel dimension to obtain long and short temporal fusion features; S414: The long and short temporal fusion features are subjected to attention enhancement processing according to the temporal attention mechanism to obtain temporal attention features.

[0022] It should be noted that, firstly, the parallel long-time-series feature extraction unit and short-time-series feature extraction unit in the temporal feature extraction enhancement module extract long-time-series and short-time-series features from the target running parameter sequence, respectively, thus obtaining the target long-time-series features and target short-time-series features. Then, the target long-time-series features and target short-time-series features are concatenated along the channel dimension to achieve fusion of the long-time-series features and target short-time-series features, resulting in long-short-time-series fused features. Finally, a temporal attention mechanism is used to enhance the attention of the long-short-time-series fused features, resulting in temporal attention features. The temporal feature extraction enhancement module extracts target long-time-series features and target short-time-series features through feature extraction by the long-time-series and short-time-series extraction units, achieves fusion of target long-time-series and target short-time-series features through concatenation, and finally processes the fused long-short-time-series fused features through a temporal attention mechanism to obtain temporal attention features. This significantly improves the feature capture capability of the temporal feature extraction enhancement module, thereby improving the reliability of network prediction.

[0023] As an optional implementation of this embodiment, in S411, the target running parameter sequence is extracted using the long-time-series feature extraction unit to obtain target long-time-series features. Specifically, S4111: The target running parameter sequence is processed using a bidirectional long short-term memory network layer to obtain a first long-time-series memory feature; S4112: The first long-time-series memory feature is enhanced by self-attention using a self-attention mechanism to obtain a long-time-series attention-enhanced feature; S4113: The long-time-series attention-enhanced feature is processed using a bidirectional long short-term memory network layer to obtain a second long-time-series memory feature; S4114: The second long-time-series memory feature is subjected to global average pooling using a pooling layer to obtain the target long-time-series feature.

[0024] It should be noted that the bidirectional Long Short-Term Memory (LSTM) network layer employs a BiLSTM network. The forward LSTM of the BiLSTM network processes the target runtime parameter sequence in chronological order, capturing left-to-right dependencies, while the backward LSTM processes the target runtime parameter sequence in reverse chronological order, capturing right-to-left dependencies. The dependencies captured by the forward and backward LSTMs are concatenated and fused at each time step to obtain the first long temporal memory feature. The first long temporal memory feature contains information from each time point and also encodes its bidirectional contextual information. Next, a self-attention mechanism enhances the first long temporal memory feature to obtain the long temporal attention-enhanced feature. The self-attention mechanism calculates the association weights between features at all time points in the first long temporal memory feature. Through this global attention calculation, the first long temporal memory feature is re-weighted and aggregated to obtain the long temporal attention-enhanced feature. Each time point in the long temporal attention-enhanced feature incorporates the data from other time points. The relevant information is then processed through a bidirectional long short-term memory (BSSM) network layer to obtain the second long-term memory feature. This BSSM layer also uses a BiLSTM network, operating on the same principle as the target parameter sequence, to obtain the second long-term memory feature. Finally, the second long-term memory feature is subjected to global average pooling through a pooling layer to obtain the target long-term feature. Global average pooling averages the value of each feature channel across the entire time dimension of the second long-term memory feature to obtain the target long-term feature. The long-term feature extraction unit inserts a self-attention mechanism between the two BSSM networks, forming an information extraction and enhancement process of temporal extraction, attention enhancement, and then temporal extraction, thereby extracting a highly continuous second long-term memory feature. This second long-term memory feature is then subjected to global average pooling to obtain the globally statistical target long-term feature. This significantly improves the network's ability to extract long-term information, thus improving the network's final prediction accuracy.

[0025] As an optional implementation of this embodiment, in S412, the target running parameter sequence is extracted using the short-time feature extraction unit to obtain target short-time features. Specifically, S4121: The target running parameter sequence is convolved using a 1×1 convolutional layer to obtain the first short-time convolutional feature; S4122: The first short-time convolutional feature is convolved using a 1×3 convolutional layer, a 1×5 convolutional layer, and a 1×7 convolutional layer to obtain the second, third, and fourth short-time convolutional features, respectively; S4123: The second, third, and fourth short-time convolutional features are convolved along... S4124: Concatenate the first short-temporal fusion feature along the channel dimension to obtain the short-temporal attention-enhanced feature; S4125: Perform attention enhancement processing on the first short-temporal fusion feature according to the self-attention mechanism to obtain the short-temporal attention-enhanced feature; S4126: Perform convolution processing on the short-temporal attention-enhanced feature according to multiple 1×3 convolutional layers to obtain the fifth short-temporal convolutional feature output by each 1×3 convolutional layer; S4127: Add the fifth short-temporal convolutional features output by each 1×3 convolutional layer element-wise to obtain the second short-temporal fusion feature; S4128: Concatenate the first short-temporal fusion feature and the second short-temporal fusion feature along the channel dimension to obtain the target short-temporal feature.

[0026] It should be noted that a 1×1 convolutional layer is used to convolve the target running parameter sequence to achieve linear combination and feature transformation of different channel features in the target parameter sequence, resulting in the first short-time convolutional feature. Then, 1×3, 1×5, and 1×7 convolutional layers are used to convolve the first short-time convolutional feature respectively, capturing information from three different receptive fields (small, medium, and large). This allows the short-time feature extraction unit to simultaneously extract local patterns at different time scales from the first short-time convolutional feature, thereby obtaining the second, third, and fourth short-time convolutional features, providing rich information for understanding complex oscillations. The system uses a rich information foundation; then, the second, third, and fourth short-temporal convolutional features are concatenated along the channel dimension to achieve a multi-resolution representation of short-temporal dynamics, thus obtaining the first short-temporal fusion feature. Following this, a self-attention mechanism is used to enhance the attention of the first short-temporal fusion feature. This mechanism maps the first short-temporal fusion feature to query features, key features, and value features. Relevance is calculated using the query and key features to obtain weighted weights, which are then multiplied by the value features to obtain the short-temporal attention-enhanced feature. The self-attention mechanism calculates the time-series data in the first short-temporal fusion feature. At the midpoint, the importance of features from different scale channels is assessed, and feature information is enhanced based on this importance. Then, multiple 1×3 convolutional layers sequentially convolve the short-temporal attention enhancement features to obtain the fifth short-temporal convolutional feature output by each of the multiple 1×3 convolutional layers. In this embodiment, five 1×3 convolutional layers sequentially convolve the short-temporal attention enhancement features, with the input of each subsequent 1×3 convolutional layer coming from the output of the previous 1×3 convolution. Thus, these five 1×3 convolutions correspond to one output, namely the fifth short-temporal convolutional feature. Successive 1×3 convolutional layers can learn more complex and higher-order fast dynamics. Then, the fifth short-time convolutional features output from each 1×3 convolutional layer are element-wise summed to obtain the second short-time fusion feature, forming a densely connected structure that ensures the effective propagation of gradients in the deep network and also achieves the fusion of features at different depth levels. Finally, the first and second short-time fusion features are concatenated along the channel dimension to achieve the fusion of the first short-time fusion feature from the multi-scale path and the second segment of the time-series fusion feature from the deep path, resulting in the target short-time feature. The target short-time feature contains rich details from multiple scales and complex patterns that have undergone deep nonlinear abstraction, making it the highest-level and most comprehensive representation of short-time information.The short-time-series feature extraction unit captures and fuses multi-scale information to obtain the first short-time-series fused feature. This first short-time-series fused feature is then subjected to multi-convolutional depth extraction, and features extracted at different depths are added and fused to obtain the second short-time-series fused feature, which is rich in information at different depths. Finally, the first short-time-series fused feature from the multi-scale path and the second short-time-series fused feature from the depth path are fused, achieving the fusion of multi-scale and depth information. This process significantly improves the network's ability to capture short-time-series information, thereby enhancing the reliability and accuracy of the network's final prediction results.

[0027] As an optional implementation of this embodiment, S414 performs attention enhancement processing on the long and short temporal fusion features according to the temporal attention mechanism to obtain temporal attention features, specifically: S4141: global average pooling processing is performed on the long and short temporal fusion features according to the pooling layer to obtain long and short time average pooling features; S4142: full connection processing is performed on the long and short time average pooling features according to the fully connected layer to obtain long and short time fully connected features; S4143: Softmax activation processing is performed on the long and short time fully connected features according to the activation layer to obtain temporal attention weights; S4144: element-wise multiplication of the temporal attention weights with the long and short temporal fusion features is performed to obtain temporal attention features.

[0028] It should be noted that, firstly, the pooling layer performs average pooling on the long-short time series fusion features, aggregating the global statistical information of each channel within the time window and calculating a scalar representing the importance of each feature channel over the entire time span, thus obtaining the long-short time average pooled features. Next, the long-short time average pooled features are processed through fully connected layers. Here, a bottleneck structure consisting of two consecutive fully connected layers is used: the first fully connected layer performs dimensionality reduction, and the second fully connected layer is used to recover the number of channels. This process allows the network to learn the non-linear interactions and dependencies between channels, resulting in long-short time fully connected features. The fully connected layer enables the network to learn the cooperative and inhibitory relationships between different feature channels. Finally, the long-short time fully connected features are processed with Softmax activation through activation layers, which transforms the learned channel importance into a probability distribution, resulting in... The process involves first determining the temporal attention weights; then, multiplying these weights element-wise with the long-short temporal fusion features to obtain the temporal attention features. These weights are then broadcast along the time dimension and multiplied with the corresponding channel at each time step of the long-short temporal fusion features. The resulting temporal attention features have the same shape as the long-short temporal fusion features, but the value of each channel is scaled and adjusted according to its importance. The temporal attention mechanism calculates the importance of each channel using global average pooling, followed by two fully connected layers for channel compression and recovery to learn the dependencies between channels. Softmax activation is then used to transform the importance into a probability distribution, which is ultimately used to modulate the long-short temporal fusion features. This process amplifies important channels while suppressing irrelevant interference channels, significantly improving the network's prediction accuracy.

[0029] S500: The control system of the target wind farm adjusts the parameters of the equipment in the target wind farm based on the prediction results.

[0030] It should be noted that the control system automatically matches and calculates the optimal control strategy based on the prediction results, namely the predicted oscillation amplitude and predicted oscillation frequency in the first time period in the future. This allows measures to be taken when the current oscillation of the target wind farm has not yet occurred or has just occurred, thereby avoiding equipment damage or system distortion caused by the current oscillation.

[0031] According to a wind farm current oscillation control method in this embodiment, the oscillation prediction of the target wind farm is achieved by acquiring the target operating parameters and sequence and the target spatial topology map of the target wind farm. Based on the prediction results, the equipment parameters of the target wind farm are adjusted in a timely manner to achieve timely and accurate prediction and suppression of the target wind farm current oscillation. The oscillation prediction network uses a dual attention mechanism consisting of a temporal and spatial feature extraction enhancement module and a spatial feature extraction enhancement module to achieve dual extraction of temporal and spatial attention features. The spatiotemporal feature fusion module fuses the temporal and spatial attention features through a cross-attention mechanism to obtain the final spatiotemporal attention features. Finally, a multi-task prediction head is used to predict the spatiotemporal attention features to achieve high-precision prediction by the oscillation prediction network. The temporal feature extraction enhancement module extracts target long-time and short-time features through long-time feature extraction units and short-time feature extraction units, respectively. The target long-time and short-time features are fused through a splicing operation. Finally, the fused long-short-time features are processed by the temporal attention mechanism to obtain the temporal attention feature. The attention feature extraction module significantly enhances the ability of the temporal feature extraction enhancement module to capture features, thereby improving the reliability of network predictions. The long temporal feature extraction unit inserts a self-attention mechanism between two bidirectional long short-term memory networks, forming an information extraction enhancement process of temporal extraction, attention enhancement, and then temporal extraction, thereby extracting a highly continuous second long temporal memory feature. Then, global average pooling is performed on the second long temporal memory feature to obtain the target long temporal feature with global statistics. This significantly improves the network's ability to extract long temporal information, thereby improving the network's final prediction accuracy. The short temporal feature extraction unit obtains the first short temporal fusion feature through multi-scale information capture and fusion, then performs multi-convolutional depth extraction on the first short temporal fusion feature, and adds and fuses the features extracted at different depths to obtain the second short temporal fusion feature rich in information of different depths. Finally, the first short temporal fusion feature from the multi-scale path and the second temporal fusion feature from the depth path are fused to achieve the fusion of multi-scale information and depth information. This process significantly improves the network's ability to capture short temporal information, thereby improving the reliability and accuracy of the network's final prediction results.

[0032] It should be noted that the wind farm current oscillation control method provided in this application embodiment can be executed by a wind farm current oscillation control system, or by a control module in the wind farm current oscillation control system for executing the wind farm current oscillation control method. This application embodiment uses an example of a wind farm current oscillation control system executing the wind farm current oscillation control method to illustrate the wind farm current oscillation control method provided in this application embodiment.

[0033] A wind farm current oscillation control system includes: a data construction module for acquiring operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset; a network construction module for constructing an oscillation prediction network, which includes a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; a network training module for training the oscillation prediction network based on a portion of the dataset and testing the trained oscillation prediction network based on another portion of the dataset; a first processing module for acquiring target operating parameter sequences and target spatial topology maps of a target wind farm, processing the target operating parameter sequences and target spatial topology maps according to the oscillation prediction network, and obtaining prediction results; and a second processing module for adjusting the parameters of the equipment in the target wind farm based on the prediction results.

[0034] The wind farm current oscillation control system in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), etc. This application embodiment does not impose specific limitations.

[0035] The wind farm current oscillation control system in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0036] The wind farm current oscillation control system provided in this application embodiment can realize the various processes and effects of the wind farm current oscillation control method in the method embodiments of Figures 1 to 3. To avoid repetition, these will not be described again here.

[0037] Optionally, this application embodiment also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described wind farm current oscillation control method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0038] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described wind farm current oscillation control method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0039] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0040] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0041] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0042] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for controlling current oscillation in a wind farm, characterized in that, The method includes: acquiring operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset based on the operating parameter sequences and spatial topology maps of multiple wind farms; constructing an oscillation prediction network, the oscillation prediction network including a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; training the oscillation prediction network based on a portion of the dataset, and testing the trained oscillation prediction network based on another portion of the dataset; acquiring a target operating parameter sequence and a target spatial topology map of a target wind farm, processing the target operating parameter sequence and the target spatial topology map according to the oscillation prediction network to obtain a prediction result; and adjusting the parameters of the equipment in the target wind farm according to the prediction result.

2. The wind farm current oscillation control method according to claim 1, characterized in that, The process of training the oscillation prediction network using a portion of the dataset and testing the trained oscillation prediction network using another portion of the dataset specifically involves: dividing the dataset into a training set and a test set according to a preset ratio, and constructing a mean squared error loss function; iteratively training the oscillation prediction network using the training set and the mean squared error loss function until the parameters of the oscillation prediction network converge; testing the converged oscillation prediction network using the test set to obtain test results, and evaluating the test results using a generalization ability evaluation index to obtain evaluation results; if the evaluation results are satisfactory, the training and testing of the oscillation prediction network are completed; if the evaluation results are unsatisfactory, the parameters of the oscillation prediction network are adjusted and retrained until the evaluation results are satisfactory.

3. The wind farm current oscillation control method according to claim 1, characterized in that, The oscillation prediction network processes the target operating parameter sequence and the target spatial topology map to obtain prediction results. Specifically, it processes the target operating parameter sequence according to the temporal feature extraction and enhancement module to obtain temporal attention features; processes the target spatial topology map according to the spatial feature extraction and enhancement module to obtain spatial attention features; fuses the temporal attention features and the spatial attention features according to the spatiotemporal feature fusion module to obtain spatiotemporal attention features; and performs prediction processing on the spatiotemporal attention features according to the multi-task prediction head to obtain the prediction results. The prediction results include the predicted oscillation amplitude and predicted oscillation frequency of the target wind farm in the first time period in the future.

4. The wind farm current oscillation control method according to claim 3, characterized in that, The temporal feature extraction and enhancement module performs temporal feature extraction processing on the target operating parameter sequence to obtain temporal attention features. Specifically, it performs long-term temporal feature extraction on the target operating parameter sequence using the long-term temporal feature extraction unit to obtain target long-term temporal features; and it performs short-term temporal feature extraction on the target operating parameter sequence using the short-term temporal feature extraction unit to obtain target short-term temporal features. The target long-time feature and the target short-time feature are concatenated along the channel dimension to obtain the long-short-time fusion feature; the long-short-time fusion feature is then subjected to attention enhancement processing according to the temporal attention mechanism to obtain the temporal attention feature.

5. The wind farm current oscillation control method according to claim 4, characterized in that, The step of extracting long-term features from the target operating parameter sequence using the long-term feature extraction unit to obtain target long-term features specifically involves processing the target operating parameter sequence using a bidirectional long short-term memory network layer to obtain a first long-term memory feature. The first long-term memory feature is subjected to self-attention enhancement processing based on the self-attention mechanism to obtain long-term attention-enhanced features; The long-term attention enhancement features are processed by a bidirectional long short-term memory network layer to obtain a second long-term memory feature; the second long-term memory feature is then subjected to global average pooling by a pooling layer to obtain the target long-term feature.

6. The wind farm current oscillation control method according to claim 4, characterized in that, The step of extracting short-time features from the target running parameter sequence using the short-time feature extraction unit to obtain target short-time features specifically involves: performing convolution processing on the target running parameter sequence using a 1×1 convolutional layer to obtain the first short-time convolutional feature. The first short-temporal convolutional feature is processed by convolutional layers of 1×3, 1×5, and 1×7 respectively to obtain the second, third, and fourth short-temporal convolutional features respectively; the second, third, and fourth short-temporal convolutional features are concatenated along the channel dimension to obtain the first short-temporal fusion feature; The first short-temporal fusion feature is subjected to attention enhancement processing based on the self-attention mechanism to obtain short-temporal attention-enhanced features; The short-temporal attention enhancement features are sequentially convolved by multiple 1×3 convolutional layers to obtain the fifth short-temporal convolutional feature output by each 1×3 convolutional layer. The fifth short-temporal convolutional features output by each of the 1×3 convolutional layers are added element-wise to obtain the second short-temporal fusion feature; The first short-time fusion feature and the second short-time fusion feature are concatenated along the channel dimension to obtain the target short-time feature.

7. The wind farm current oscillation control method according to claim 4, characterized in that, The process of enhancing the attention of the long-short temporal fusion features according to the temporal attention mechanism to obtain the temporal attention features specifically involves: performing global average pooling on the long-short temporal fusion features using a pooling layer to obtain long-short temporal average pooling features; performing full connection processing on the long-short temporal average pooling features using a fully connected layer to obtain long-short temporal fully connected features; performing Softmax activation processing on the long-short temporal fully connected features using an activation layer to obtain temporal attention weights; and performing element-wise multiplication of the temporal attention weights with the long-short temporal fusion features to obtain the temporal attention features.

8. A wind farm current oscillation control system, capable of implementing the wind farm current oscillation control method according to any one of claims 1-7, characterized in that, The system includes: a data construction module: acquiring operating parameter sequences and spatial topology maps of multiple wind farms to construct a dataset based on these sequences; a network construction module: constructing an oscillation prediction network, which includes a temporal feature extraction enhancement module, a spatial feature extraction enhancement module, a spatiotemporal feature fusion module, and a multi-task prediction head; a network training module: training the oscillation prediction network based on a portion of the dataset and testing the trained network based on another portion of the dataset; a first processing module: acquiring the target operating parameter sequence and target spatial topology map of the target wind farm, processing the sequence and map using the oscillation prediction network to obtain prediction results; and a second processing module: adjusting the parameters of the equipment at the target wind farm based on the prediction results, according to the prediction results.

9. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of a wind farm current oscillation control method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of a wind farm current oscillation control method as described in any one of claims 1-7.

Images