Oscillation source positioning method and device for large-scale wind power cluster based on deep learning

Abstract

The application discloses a large-scale wind power cluster oscillation source positioning method and device based on deep learning, and the method comprises the following steps: acquiring PMU data and state estimation data based on PMU of each wind power cluster as input data; converting the input data into a two-dimensional image; constructing a sample set and a sub-sample set according to the two-dimensional image; constructing and training a layered convolutional neural network with coarse granularity and fine granularity according to the sample set and the sub-sample set, and positioning the wind power cluster oscillation source. The application considers the wind power grid-connected system oscillation instability phenomenon caused by dynamic interaction among multiple wind power clusters, can identify and accurately position multiple oscillation sources, and is suitable for complex and changeable operation scenes of the wind power grid-connected system.

Description

Technical Field

[0001] This invention relates to wind power grid-connected stability analysis and control technology, particularly to large-scale wind power cluster oscillation source localization technology, specifically to a method and apparatus for large-scale wind power cluster oscillation source localization based on deep learning. Background Technology

[0002] In recent years, the proportion of wind power generation has continued to grow, and wind power has gradually become an important part of the energy structure. Wind power bases and load centers exhibit an inverse distribution pattern; in addition to local consumption, wind power from the northwest region is transmitted on a large scale and over long distances to East China, Central China, and North China, forming a power grid pattern of west-to-east and north-to-south power transmission. However, wind power grid connection has brought many adverse effects and shocks to the safe and stable operation of the power system.

[0003] Disconnecting all wind farm clusters during an oscillation instability incident can ensure the safety and stability of the power system, but at an excessive cost. If the oscillation source within the wind farm cluster is quickly and promptly disconnected at the incident site, power generation losses can be minimized, balancing safety and economy. This relies on the precise location of the oscillation source. However, the dynamic interaction process of wind power grid-connected systems is complex, and the oscillation instability mechanism is not yet fully understood. Mechanism-based methods for locating oscillation sources in wind farm clusters have inherent limitations. A promising technical approach is measurement data-driven methods. The current technical bottleneck lies in the fact that large-scale wind farms typically contain dozens or even hundreds of wind farm clusters, and the system operation scenarios are complex and variable. Existing measurement data-driven oscillation source location technologies are insufficient to meet the requirements for precise location. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method and device for locating oscillation sources in large-scale wind power clusters based on deep learning. This solution considers the oscillation and instability problem of wind power grid-connected systems caused by dynamic interactions between multiple wind power clusters, and can identify and accurately locate multiple oscillation sources, adapting to the complex and ever-changing operating scenarios of wind power grid-connected systems.

[0005] Technical solution: The deep learning-based method for locating oscillation sources in large-scale wind power clusters of the present invention includes the following steps:

[0006] (1) Obtain PMU data and PMU-based state estimation data for each wind power cluster as input data;

[0007] (2) Convert the input data into a two-dimensional image;

[0008] (3) Construct a sample set and a sub-sample set based on the two-dimensional image;

[0009] (4) Construct and train coarse-grained and fine-grained hierarchical convolutional neural networks based on the sample set and sub-sample set to locate the oscillation source of the wind power cluster.

[0010] In step (1), the PMU data and state estimation data based on the PMU for each wind power cluster include the port voltage amplitude V of all N wind power clusters under M power system operating conditions. ij Wind power cluster port voltage phase T ij The active power output P of the wind power cluster ij And the reactive power output Q of the wind power cluster ij The time series, where i = 1, 2, ..., M, j = 1, 2, ..., N.

[0011] In step (2), converting the input data into a two-dimensional image includes converting the input data V, which contains time-series features, based on the Gram angle field. ij T ij P ij Q ij Convert each image into a two-dimensional image, and represent all V values ​​of the N wind power clusters under operating condition i. ij T ij P ij Q ij The images are stitched together to form a single image, resulting in M ​​two-dimensional images for M different power system operating conditions.

[0012] In step (3), the construction of the sample set and sub-sample set includes the following steps:

[0013] (3.1) Using two-dimensional images as sample features and the number of the oscillation-dominant wind power cluster as sample label, a sample set is constructed;

[0014] (3.2) Based on the division of wind power clusters as the dominant oscillation / participation in dynamic interaction, the sample set is divided into N sub-sample sets; each sub-sample set i consists of samples of wind power cluster i as the dominant oscillation or participation in dynamic interaction, and the sub-sample set is labeled with the participation degree of each wind power cluster.

[0015] In step (4), the specific steps for constructing and training coarse-grained and fine-grained hierarchical convolutional neural networks based on the sample set and sub-sample set to locate the oscillation source of the wind power cluster are as follows:

[0016] (4.1) Construct and train a coarse-grained convolutional neural network based on the sample set to achieve preliminary localization of the dominant oscillation source;

[0017] (4.2) Construct and train N fine-grained convolutional neural networks based on N subsets of samples. The fine-grained convolutional neural networks are used to determine the degree of oscillation participation of each wind power cluster and locate multiple oscillation sources that are dynamically interacting.

[0018] (4.3) The output of the measurement data through the coarse-grained convolutional neural network is the number of the oscillation-dominant wind power cluster. The i-th fine-grained convolutional neural network is selected according to the output wind power cluster number i.

[0019] (4.4) The output of the measurement data through the i-th fine-grained convolutional neural network is the degree of oscillation participation of each wind power cluster. The top h wind power clusters with the highest participation are identified as oscillation sources.

[0020] The present invention also includes a deep learning-based device for locating oscillation sources in large-scale wind power clusters, comprising a data preprocessing module for acquiring PMU data and state estimation data based on the PMU of the wind power clusters and transmitting them to a data conversion module; the data conversion module for receiving the output data from the data preprocessing module, converting it into a two-dimensional image, and transmitting it to a deep learning computing module; and the deep learning computing module for storing coarse-grained and fine-grained hierarchical convolutional neural network model parameters, and performing model calculations based on the received two-dimensional image output by the data conversion module to locate the oscillation sources in the wind power clusters and output the oscillation participation level of each wind power cluster.

[0021] The data preprocessing module stores and loads the port voltage amplitude V of all N wind power clusters under the current operating conditions. 0k Wind power cluster port voltage phase T 0k The active power output P of the wind power cluster 0k The reactive power output Q of the wind power cluster 0k The time series, k = 1, 2, ..., N.

[0022] The data conversion module converts the time-series data from the data preprocessing module into a two-dimensional image, including, based on the Gram angle field algorithm, converting the time-series feature data V... 0k T 0k P 0k Q 0k Convert each into a two-dimensional image, and then convert all V values ​​of the N wind power clusters. 0k T 0k P 0k Q 0k Combined into a single image.

[0023] The deep learning computing module stores coarse-grained and fine-grained hierarchical convolutional neural networks trained offline. Based on the two-dimensional image generated by the data conversion module, the coarse-grained convolutional neural network model calculates and outputs the number of the dominant oscillation wind power cluster. Based on the output wind power cluster number i, the i-th fine-grained convolutional neural network is selected. Based on the two-dimensional image generated by the data conversion module, the i-th fine-grained convolutional neural network model calculates and outputs the oscillation participation degree of each wind power cluster, and the top h wind power clusters with the highest participation degree are identified as oscillation sources.

[0024] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) It takes into account the multi-source oscillation instability caused by the dynamic interaction between multiple wind power clusters, and adopts coarse-grained and fine-grained hierarchical convolutional neural network to enhance the adaptability of the model to complex and variable operating scenarios, improves the accuracy of oscillation source positioning, helps to quickly and timely cut off the oscillation source of the wind power cluster in the accident site, minimizes power generation loss, and has significant economic benefits; (2) It is applicable to the precise optimization control of wind power clusters at the station level and the emergency control of the power grid, ensuring the safe and stable operation of the power grid. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process of the present invention;

[0026] Figure 2 This is a schematic diagram of the large-scale wind power grid connection system in this invention;

[0027] Figure 3 This is a schematic diagram of an example power system structure containing 5 wind power clusters in this invention;

[0028] Figure 4 This is a schematic diagram illustrating the conversion of a time series into a two-dimensional image based on Gram angle field in this invention;

[0029] Figure 5 This is a schematic diagram illustrating the conversion of all wind power cluster time series data into a two-dimensional image based on the Gram angle field in this invention;

[0030] Figure 6 This is a diagram of the convolutional neural network structure in this invention;

[0031] Figure 7 This is a diagram showing the coarse-grained and fine-grained hierarchical convolutional neural network structures in this invention. Detailed Implementation

[0032] The technical solution of the present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0033] like Figure 1 and Figure 2 As shown, for various operating scenarios of wind power grid-connected systems, this invention proposes a method for locating oscillation sources in large-scale wind power clusters based on deep learning, including the following steps:

[0034] (1) Obtain PMU data and state estimation data based on PMU for each wind power cluster as input data, including the port voltage amplitude V of all N wind power clusters under M power system operating conditions. ij Wind power cluster port voltage phase T ij The active power output P of the wind power cluster ijAnd the reactive power output Q of the wind power cluster ij The time series, where i = 1, 2, ..., M, j = 1, 2, ..., N;

[0035] (2) Convert the input data into a two-dimensional image: Based on the Gram angle field, convert the input data V with time series features into a two-dimensional image. ij T ij P ij Q ij Convert each image into a two-dimensional image, and represent all V values ​​of the N wind power clusters under operating condition i. ij T ij P ij Q ij The images are stitched together to form a single image, resulting in M ​​two-dimensional images for M different power system operating conditions.

[0036] (3) Construct a sample set and a sub-sample set based on the two-dimensional image, including the following steps:

[0037] (3.1) Construct a sample set using all two-dimensional images as sample features and the number of the oscillation-dominant wind power cluster as the sample label;

[0038] (3.2) Based on the division of wind power clusters as the dominant oscillation / participation in dynamic interaction, the sample set is divided into N sub-sample sets; each sub-sample set i consists of samples of wind power cluster i as the dominant oscillation or participation in dynamic interaction, and the sub-sample set is labeled with the participation degree of each wind power cluster.

[0039] (4) Construct and train coarse-grained and fine-grained hierarchical convolutional neural networks based on the sample set and sub-sample set to locate the oscillation source of the wind power cluster. The specific steps are as follows:

[0040] (4.1) Construct and train a coarse-grained convolutional neural network based on the sample set to achieve preliminary localization of the dominant oscillation source;

[0041] (4.2) Construct and train N fine-grained convolutional neural networks based on N subsets of samples. The fine-grained convolutional neural networks are used to determine the degree of oscillation participation of each wind power cluster and locate multiple oscillation sources that are dynamically interacting.

[0042] (4.3) The output of the measurement data through the coarse-grained convolutional neural network is the number of the oscillation-dominant wind power cluster. The i-th fine-grained convolutional neural network is selected according to the output wind power cluster number i.

[0043] (4.4) The output of the measurement data through the i-th fine-grained convolutional neural network is the degree of oscillation participation of each wind power cluster. The top h wind power clusters with the highest participation are identified as oscillation sources.

[0044] This invention takes into account the oscillation and instability phenomenon of wind power grid-connected systems caused by dynamic interaction between multiple wind power clusters, and can identify multiple oscillation sources, adapting to the complex and ever-changing operating scenarios of wind power grid-connected systems.

[0045] Example 1

[0046] like Figure 3 As shown, stability analysis of the system under actual operating conditions is performed in the simulation system, and training samples are obtained to establish an oscillation source localization model for a wind power grid-connected system containing five wind turbine clusters. Specifically, based on the operating mode of the actual system, the operating point in the simulation system is set to ±20% relative to the actual system operating point: the rated power of each unit is increased or decreased by 0-20% proportionally, and the load on the line is set to 80%-120% of the actual system operating load. Then, operating points are randomly selected, and the control parameters of the wind turbine clusters are randomly set to simulate different equivalent wind turbine parameters of the actual system.

[0047] Time-domain simulations were performed based on the disturbances and frequency settings of the actual system, and measurement data X from five wind power cluster ports were recorded. s This includes the port voltage amplitude V of five wind power clusters. ij Port voltage phase T ij Output active power P ij and output reactive power Q ij The constructed time series is used to calculate the participation factors using the state-space equations of the parameterized model of the simulation system, and the sample data obtained from each simulation are labeled to form the true label set Y of the simulation system. s Thus, a large number of labeled training samples are generated in the simulation system.

[0048] The Gram angle field algorithm is used to convert time-series measurement data into two-dimensional images, which involves the following steps:

[0049] (a) Use segmented aggregation to approximate the aggregation of time series data so that the time series maintains an appropriate length;

[0050] (b) Normalize the data to obtain the scaling value for the interval [0,1];

[0051] (c) Using the timestamp as the radius and the inverse cosine of the normalized data corresponding to each timestamp as the angle, generate polar coordinates;

[0052] (d) Add up each angle value and take the cosine value to generate the Gram angle field.

[0053] like Figure 4As shown, taking the voltage amplitude time series data of wind power cluster 1 as an example, the Gram angle field is generated according to the above steps. The voltage amplitude, voltage phase, output active power, and output reactive power of wind power cluster 1 are stitched together in a row. The voltage amplitude, voltage phase, output active power, and output reactive power of wind power clusters 1 to 5 are then stitched together column by column, as shown. Figure 5 As shown, this constitutes a two-dimensional image under this operating condition. M two-dimensional images are obtained for M different operating conditions, thereby generating sample features in image form.

[0054] Using convolutional neural networks (CNNs) to locate oscillation sources avoids issues such as sampling frequency leakage caused by upload speed. By inputting low-redundancy data for oscillation source localization, not only is the number of parameters required to train the CNN model reduced, improving the localization model's performance, but it also demonstrates that the data processed by the CNN retains the key features of the oscillations.

[0055] Convolutional neural networks (CNNs) are a special type of multilayer perceptron, initially proposed for learning two-dimensional shapes (images). With the introduction of convolution and pooling, CNNs have demonstrated excellent feature extraction capabilities, making them suitable for various fields. Since system dynamics can be well described using two-dimensional images, for problems like broadband oscillations where the mechanism is unclear and the features are difficult to express definitively, CNNs can be used as the main network framework to achieve feature extraction. Figure 6 The diagram shows a typical convolutional neural network structure, consisting of an input layer, convolutional layers, pooling layers, ReLU, and fully connected layers.

[0056] In convolutional neural networks, convolutional layers extract local features from the input data through convolution operations. A convolutional layer consists of multiple feature matrices, which are convolved with a convolutional kernel to obtain a new feature matrix for the next layer. The calculation formula is as follows:

[0057]

[0058] Among them, c mn x is the element in the m-th row and n-th column of the output matrix c; m+i，n+j Let w be the element in the (m+i)th row and (n+j)th column of the input matrix x; ij is the element in the i-th row and j-th column of the k-th convolution kernel w; b is the bias term; f is the activation function.

[0059] The pooling layer reduces the dimensionality of the feature matrix. Similar to filtering, pooling reduces the dimensionality of the data, preserving key features while decreasing computational cost. It also prevents overfitting. Furthermore, this process is parameter-free, eliminating the need for parameter updates. This paper employs mean pooling, calculated using the following formula:

[0060]

[0061] In the formula: y ab c is the element in the a-th row and b-th column of the output matrix y; a·p+i，b·p+j Let be the elements in the a·p+i-th row and b·p+j-th column of the input matrix c.

[0062] Pooling layers are typically placed after convolutional layers. A multi-layered convolutional-pooling network structure can achieve good feature recognition results for deep convolutional networks. The extracted features are unfolded into a one-dimensional array and input into a fully connected layer. After being mapped by a ReLU function, the output yields the measured feature representation.

[0063] A convolutional neural network (CNN) is constructed, using sample features in the form of two-dimensional images and the ID numbers of oscillating wind power clusters as sample labels. The CNN model parameters are trained using a stochastic gradient descent algorithm. However, the operating scenarios of wind power grid-connected systems are complex and varied, and a single CNN can hardly accurately handle all situations. Therefore, in this invention, the CNN is referred to as a coarse-grained CNN. To improve the accuracy of oscillation source localization, a fine-grained CNN is constructed as the next layer after the coarse-grained CNN.

[0064] The training sample set for the coarse-grained convolutional neural network is divided into five subsets. The generation principle for these subsets is as follows: based on the participation factor, each sample generates oscillation participation data for each wind power cluster. If wind power cluster i has a high participation rate, the sample belongs to subset i. Obviously, due to the mechanism of oscillation instability caused by multi-source dynamic interaction, some samples may contain multiple wind power clusters with high participation factors; therefore, some samples may belong to multiple subsets simultaneously. Thus, the sample features in the subsets remain two-dimensional images, and the sample label is the oscillation participation rate of each wind power cluster. Therefore, the i-th fine-grained convolutional neural network is trained based on the i-th subset. In practice, the obtained measurement data is processed by the coarse-grained convolutional neural network to calculate and output the number of the dominant oscillating wind power cluster. The i-th fine-grained convolutional neural network is selected based on the output wind power cluster number i. Then, the measurement data is processed by the i-th fine-grained convolutional neural network to calculate and output the oscillation participation rate of each wind power cluster, and the top h wind power clusters with the highest participation rates are identified as oscillation sources.

[0065] Therefore, coarse-grained and fine-grained hierarchical convolutional neural network models were constructed, and the model architecture is as follows: Figure 7 As shown, it can achieve precise positioning of multiple oscillation sources.

[0066] Example 2

[0067] Based on the same inventive concept as Embodiment 1, this embodiment introduces a deep learning-based oscillation source localization device for large-scale wind power clusters, comprising:

[0068] The data preprocessing module is configured to acquire PMU data and PMU-based state estimation data of the wind power cluster; the data preprocessing module stores and loads the port voltage amplitude V of all N wind power clusters under the current operating conditions. 0k Wind power cluster port voltage phase T 0k The active power output P of the wind power cluster 0k And the reactive power output Q of the wind power cluster 0k The time series, k = 1, 2, ..., N.

[0069] The data conversion module is configured to convert the PMU data and PMU-based state estimation data of the wind power cluster into a two-dimensional image; the data conversion module converts the time series data from the data preprocessing module into a two-dimensional image, including converting the time series feature data V based on the Gram angle field algorithm. 0k T 0k P 0k Q 0k Convert each into a two-dimensional image, and then convert all V values ​​of the N wind power clusters. 0k T 0k P 0k Q 0k Combined into a single image.

[0070] The deep learning computing module is configured to store parameters of coarse-grained and fine-grained hierarchical convolutional neural network models and perform model calculations to locate the oscillation sources of wind power clusters. The deep learning computing module also stores offline-trained coarse-grained and fine-grained hierarchical convolutional neural networks. Based on the two-dimensional image generated by the data conversion module, the coarse-grained convolutional neural network model calculates and outputs the number of the dominant oscillating wind power cluster. Based on the output wind power cluster number i, the i-th fine-grained convolutional neural network is selected. Based on the two-dimensional image generated by the data conversion module, the i-th fine-grained convolutional neural network model again calculates and outputs the oscillation participation degree of each wind power cluster, and the top h wind power clusters with the highest participation degree are identified as oscillation sources.

Claims

1. A method for locating oscillation sources in large-scale wind power clusters based on deep learning, characterized in that, Includes the following steps: (1) Obtain PMU data and PMU-based state estimation data for each wind power cluster as input data; (2) Convert the input data into a two-dimensional image; (3) Construct a sample set and a sub-sample set based on the two-dimensional image; (4) Construct and train coarse-grained and fine-grained hierarchical convolutional neural networks based on the sample set and sub-sample set to locate the oscillation source of the wind power cluster; In step (4), the specific steps for constructing and training coarse-grained and fine-grained hierarchical convolutional neural networks based on the sample set and sub-sample set to locate the oscillation source of the wind power cluster are as follows: (4.1) Construct and train a coarse-grained convolutional neural network based on the sample set to achieve preliminary localization of the dominant oscillation source; (4.2) Construct and train N fine-grained convolutional neural networks based on N subsets of samples. The fine-grained convolutional neural networks are used to determine the degree of oscillation participation of each wind power cluster and locate multiple oscillation sources that are dynamically interacting. (4.3) The output of the measurement data through the coarse-grained convolutional neural network is the number of the oscillation-dominant wind power cluster. The i-th fine-grained convolutional neural network is selected according to the output wind power cluster number i. (4.4) The output of the measurement data through the i-th fine-grained convolutional neural network is the degree of oscillation participation of each wind power cluster. The top h wind power clusters with the highest participation are identified as oscillation sources.

2. The method for locating oscillation sources in large-scale wind power clusters based on deep learning according to claim 1, characterized in that: In step (1), the PMU data and state estimation data based on the PMU for each wind power cluster include the port voltage amplitude V of all N wind power clusters under M power system operating conditions. ij Wind power cluster port voltage phase T ij The active power output P of the wind power cluster ij The reactive power output Q of the wind power cluster ij The time series, where i = 1, 2, ..., M, j = 1, 2, ..., N.

3. The method for locating oscillation sources in large-scale wind power clusters based on deep learning according to claim 2, characterized in that: In step (2), converting the input data into a two-dimensional image includes converting the input data V, which contains time-series features, based on the Gram angle field. ij T ij P ij Q ij Convert each image into a two-dimensional image, and represent all V values ​​of the N wind power clusters under operating condition i. ij T ij P ij Q ij The images are stitched together to form a single image, resulting in M ​​two-dimensional images for M different power system operating conditions.

4. The method for locating oscillation sources in large-scale wind power clusters based on deep learning according to claim 1, characterized in that, In step (3), the construction of the sample set and sub-sample set includes the following steps: (3.1) Using two-dimensional images as sample features and the number of the oscillation-dominant wind power cluster as sample label, a sample set is constructed; (3.2) Based on the division of wind power clusters as the dominant oscillation / participation in dynamic interaction, the sample set is divided into N sub-sample sets; each sub-sample set i consists of samples of wind power cluster i as the dominant oscillation or participation in dynamic interaction, and the sub-sample set is labeled with the participation degree of each wind power cluster.

5. A deep learning-based oscillation source localization device for large-scale wind power clusters, characterized in that: The method for locating oscillation sources in large-scale wind power clusters based on deep learning, as described in claim 1, includes: The data preprocessing module is used to acquire PMU data and state estimation data based on PMU of the wind power cluster, and transmit them to the data conversion module; The data conversion module receives the output data from the data preprocessing module, converts it into a two-dimensional image, and transmits it to the deep learning computing module. The deep learning computing module stores parameters of coarse-grained and fine-grained hierarchical convolutional neural network models, performs model calculations based on the two-dimensional images output by the received data conversion module, locates the oscillation source of the wind power cluster, and outputs the oscillation participation level of each wind power cluster.

6. The deep learning-based oscillation source localization device for large-scale wind power clusters according to claim 5, characterized in that: The data preprocessing module stores and loads the port voltage amplitude V of all N wind power clusters under the current operating conditions. 0k Wind power cluster port voltage phase T 0k The active power output P of the wind power cluster 0k The reactive power output Q of the wind power cluster 0k The time series, k = 1, 2, ..., N.

7. The deep learning-based oscillation source localization device for large-scale wind power clusters according to claim 6, characterized in that: The data conversion module converts the time-series data from the data preprocessing module into a two-dimensional image, including, based on the Gram angle field algorithm, converting the time-series feature data V... 0k T 0k P 0k Q 0k Convert each into a two-dimensional image, and then convert all V values ​​of the N wind power clusters. 0k T 0k P 0k Q 0k Combined into a single image.

8. The deep learning-based oscillation source localization device for large-scale wind power clusters according to claim 6, characterized in that: The deep learning computing module stores coarse-grained and fine-grained hierarchical convolutional neural networks that have been trained offline. Based on the two-dimensional image generated by the data conversion module, the number of the oscillation-dominant wind power cluster is calculated and output by a coarse-grained convolutional neural network model. Select the i-th fine-grained convolutional neural network based on the output wind power cluster number i; Based on the two-dimensional image generated by the data conversion module, the oscillation participation degree of each wind power cluster is calculated and output again by the i-th fine-grained convolutional neural network model, and the top h wind power clusters with the highest participation degree are identified as oscillation sources.

Images