A method for predicting low wind speed events of wind turbines based on graph convolutional neural network

Abstract

The present invention relates to a method for predicting low wind speed events of wind turbines based on graph convolutional neural networks. By introducing an attribute graph representing the attribute similarity between geographic units and a spatial graph of spatial correlation, and combining graph convolutional neural networks and graph attention neural networks, two deep learning technologies focusing on graph structure, the multi-source geographic element characteristics and historical wind speed characteristics of geographic units with similar attributes and spatial correlation are fully utilized, so as to better understand the influence of the attribute relationship and spatial relationship between geographic units on the model prediction of future wind speed, and improve the accuracy of wind speed prediction.

Description

Technical Field

[0001] The present invention belongs to the field of computers, and in particular relates to a method for predicting low wind speed events of wind turbines based on a graph convolutional neural network. Background Art

[0002] In recent years, the growth of electricity demand has made renewable energy, including wind power, increasingly important. Wind speed changes are highly random and intermittent, and this instability can seriously affect the safety of the power system and pose a severe challenge to timely and reliable wind speed forecasting.

[0003] Traditional wind speed prediction methods include physical methods, statistical methods, and machine learning methods. Physical methods mainly predict wind speed by using physical parameters such as meteorological factors and geographical factors. However, the physical model has high computational cost and cannot capture the complex dynamic change relationship of meteorological factors, so it is not suitable for short-term wind speed prediction. Statistical methods use the linear relationship between various variables in historical time series to construct statistical models, such as time series methods, autoregressive moving average (ARMA) methods, and Kalman filtering methods. These methods overcome the shortcomings of physical models, but they can only analyze the linear relationship between variables in historical time series and have difficulty in dealing with nonlinear relationships between meteorological elements. Machine learning such as support vector machines (SVM), multilayer perceptrons (MLP), and extreme learning machines (ELM) can extract complex nonlinear features in wind speed time series and improve the accuracy of forecasts to a certain extent. However, these traditional linear and nonlinear models can only extract shallow features and require a lot of feature engineering, and cannot automatically extract time series features.

[0004] Unlike machine learning methods, deep learning uses a distributed hierarchical feature representation method to automatically extract abstract features and hidden invariant structures inherent in the data from the lowest to the highest level. Wang et al. proposed a convolutional neural network model based on an ensemble method to achieve probabilistic wind power forecasting, and verified the accuracy and stability of the model through case studies. Ghaderi et al. used a long short-term memory network model based on STC for wind speed forecasting. In addition, a recent study showed that there is a significant cross-correlation between the wind speed of the target geographic unit and its neighboring geographic units. Therefore, making full use of the meteorological information of the target geographic unit and its surrounding geographic units to improve the accuracy and reliability of wind speed forecasts has become a research hotspot. Yuan Mimi et al. proposed a CNN-LSTM model with a multi-input single-output combined structure, which can utilize the spatial feature extraction capability of CNN and the time series feature extraction capability of LSTM to further improve the effect of wind speed forecasting.

[0005] The prior art has the following disadvantages:

[0006] (1) The influence of the geographical element characteristics and wind speed characteristics of geographical units with similar attributes on the predicted wind speed of the target geographical unit is not fully considered: The third law of geography states that "the more similar the geographical configurations of two points (regions), the more similar the values ​​(processes) of the target variable at these two points (regions)". The present invention applies this law to the task of predicting the wind speed of geographical units, that is, the more similar the geographical element attributes of two geographical units are, the more similar the values ​​(processes) of the wind speed are. The wind speed of one geographical unit may not directly affect the wind speed of another geographical unit. However, if one of the geographical units has shown a certain wind speed fluctuation in a time series, according to the third law of geography, the present invention believes that when the model predicts the wind speed of another geographical unit, it will be affected by the known wind speed fluctuation of the geographical unit with similar attributes, while the prior art does not model this influence. The present invention uses a graph convolution module to aggregate the geographical element characteristics and wind speed characteristics of geographical units with similar attributes in a past time series to achieve feature extraction, obtains the wind speed prediction results of a future time series through forward propagation of a neural network, calculates the loss based on the prediction results and the actual wind speed, and completes the model training by backpropagating the error layer by layer. The geographical element characteristics and wind speed characteristics of geographical units with similar attribute relationships in a past time series and the wind speed in a future time series will participate in the model training process, aiming to enable the model to capture the impact of the wind speed conditions that have appeared in the past time series on the future wind speed of geographical units with similar attribute relationships.

[0007] (2) Ignoring spatial heterogeneity, and not fully considering the impact of the geographical element characteristics of spatially related geographical units on the predicted wind speed of the target geographical unit: The first law of geography states that "everything is related to everything else, but near things are more related than distant things". The second law of geography originated from Anselin

[18] The law of "spatial heterogeneity" is proposed. This law summarizes another characteristic of geographical phenomena: the spatial variation of geographical phenomena and the differences in variation, that is, the uncontrollable law of spatial variation. The present invention interprets this law in the task of predicting the wind speed of a geographical unit as follows: when predicting the wind speed of a geographical unit, it is necessary not only to predict based on its own geographical elements and historical wind speed, but also to consider the geographical elements and historical wind speed of adjacent geographical units. Due to the existence of spatial heterogeneity, the spatial autocorrelation between geographical units in different regions and adjacent geographical units is also different. However, the prior art uses shared weights for spatial convolution, but ignores the important influence of spatial heterogeneity between regions on the prediction of the wind speed of the target geographical unit, resulting in the accuracy of the prediction model being restricted. Summary of the invention

[0008] The purpose of the present invention is to overcome the above shortcomings of the prior art. By introducing an attribute graph representing the attribute similarity between geographic units and a spatial graph of spatial correlation, and combining two deep learning technologies focusing on graph structure, namely, graph convolutional neural network (GCN) and graph attention neural network (GAT), the multi-source geographic element characteristics and historical wind speed characteristics of geographic units with similar attributes and spatial correlation are fully utilized, so as to better understand the influence of the attribute relationship and spatial relationship between geographic units on the model prediction of future wind speed, and improve the accuracy of wind speed prediction.

[0009] Based on the similarity of multi-source geographical elements that affect the occurrence of low wind speed events, the present invention studies a method of modeling attribute graphs between geographical units and capturing wind speed characteristics of geographical units with similar attributes with the help of graph convolutional neural networks.

[0010] The present invention studies a method for modeling the spatial dependency between geographic units based on the spatial autocorrelation described in the first law of geography and the spatial heterogeneity described in the second law of geography, and mining the wind speed correlation between geographic units with the help of a graph attention mechanism.

[0011] The present invention proposes a method for predicting low wind speed events of wind turbines based on graph convolutional neural networks, which specifically includes:

[0012] (1) Construction of geographic unit attribute graph based on similarity of multi-source geographic elements: The present invention takes a single geographic unit as a node in the graph model, and takes the multi-source geographic elements that affect the occurrence of low wind speed events as the attribute information of the geographic unit. The present invention measures the attribute similarity between geographic units in a certain time series as the weight of the edge between nodes. The higher the similarity, the higher the weight, thereby completing the construction of the attribute graph. The attribute graph describes the attribute similarity between geographic units without considering the distance between the geographic units in Euclidean space. The present invention uses the attribute graph as an adjacency matrix for subsequent graph convolution modules, which can effectively capture the wind speed characteristics of geographic units with similar attributes to enhance the prediction ability of the wind speed trend of the target geographic unit.

[0013] (2) Construction of a spatial graph of geographic units based on Euclidean spatial distance: The present invention uses a single geographic unit as a node in the graph model. The present invention sets a distance threshold to measure whether there is spatial correlation between geographic units. If the Euclidean spatial distance between two geographic units is greater than the set threshold, the spatial correlation between the other geographic unit and it is not considered when predicting the wind speed trend of one of the geographic units. Edges are established between geographic unit nodes whose Euclidean spatial distance is less than the set threshold, and the weights of the edges are set as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for the subsequent graph attention module, and can effectively aggregate the geographic elements and wind speed characteristics of adjacent geographic units according to the weights to enhance the prediction ability of the wind speed trend of the target geographic unit.

[0014] (3) Graph neural network modeling for wind speed prediction: Graph neural network (GNN) is a deep learning technology that focuses on graph structure. It can aggregate information of related regions and help simulate complex associations in geographic space. Graph convolutional neural network (GCN) and graph attention neural network (GAT) are both types of graph neural networks. The present invention establishes a graph neural network model for wind speed prediction, uses the attribute graph as the adjacency matrix for the graph convolution module, uses the spatial graph as the adjacency matrix for the graph attention module, and convolves the time dimension to achieve end-to-end wind speed time series prediction.

[0015] (4) Sample dataset construction: It is necessary to construct a time series sample dataset of geographical units including air pressure (Pa), ambient temperature (°C), dew point temperature (°C), zonal wind speed at 10 meters (m / s), and meridional wind speed at 10 meters (m / s) for model training, verification, and testing. These datasets are key resources in the research and development process.

[0016] Furthermore, the present invention proposes a method for predicting low wind speed events of wind turbines based on a graph convolutional neural network, and the method specifically comprises the following steps:

[0017] Step 1: construct a geographic unit attribute graph based on the similarity of multi-source geographic elements; use a single geographic unit as a node in the graph model, and use the multi-source geographic elements that affect the occurrence of low wind speed events as the attribute information of the geographic unit; measure the attribute similarity between geographic units in a certain time series as the weight of the edge between nodes, the higher the similarity, the higher the weight, thereby completing the construction of the attribute graph;

[0018] Step 2: construct a spatial graph of geographic units based on Euclidean spatial distance; use a single geographic unit as a node in the graph model, set a distance threshold to measure whether there is spatial correlation between geographic units, and if the Euclidean spatial distance between two geographic units is greater than the set threshold, then when predicting the wind speed trend of one of the geographic units, the spatial correlation between the other geographic unit and it is not considered, and establish edges between the geographic unit nodes whose Euclidean spatial distance is less than the set threshold, and set the weight of the edge as a learnable parameter to form a spatial graph;

[0019] Step 3: Modeling the wind speed prediction graph neural network; using the attribute graph as the adjacency matrix for the graph convolution module, using the spatial graph as the adjacency matrix for the graph attention module, and convolving the time dimension to achieve end-to-end wind speed time series prediction;

[0020] Step 4: Sample data set construction and model optimization: construct a time series sample data set of geographic units including air pressure, ambient temperature, dew point temperature, zonal wind speed at a height of 10 meters, and meridional wind speed at a height of 10 meters for model training, verification, and testing.

[0021] Preferably, in step 1, the multi-source geographical elements that affect the occurrence of low wind speed events, including air pressure, ambient temperature, dew point temperature, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, slope, aspect, elevation and surface cover type, a total of 9 dimensional geographical elements are used as attribute information of the geographical unit; the wind speed prediction target area Divide in space into square geographic units, the resolution of the geographic unit is m 2 ;

[0022] By plot composition; set The areas are , the temperature values ​​are , set geographic unit The area is , the temperature value is , the calculation method is shown in formula (1):

[0023] (1)

[0024] The geographical unit The dimensional feature map is a node dimensional features, assuming that there are a total of geographical units in the region indivual, , to measure the geographical units With geographical units Taking the attribute similarity between two dimensions as an example, we measure the similarity of each dimension of geographical elements in a certain time series. The similarity of, and sum, and then find the average value, as a geographical unit With geographical units The similarity of attributes between , as shown in formula (2):

[0025] (2)

[0026] In the formula, Represents the dimension of geographic features, To record geographical units exist Dimensions a vector of sequence values, , Record geographical units With geographical units exist Dimensions a vector of sequence values, is the covariance between two vectors, is the variance of a vector;

[0027] Calculate attribute similarity between geographic units separately As the weight of the edge between nodes, build the adjacency matrix of the attribute graph :

[0028] .

[0029] Preferably, in step 2, a distance threshold is set To measure whether there is spatial correlation between geographic units, if the Euclidean distance between two geographic units Greater than the set threshold , then when predicting the wind speed trend of one geographical unit, the spatial correlation between the other geographical unit and it is not considered; in the Euclidean space distance Less than the set threshold Establish edges between geographic unit nodes and set the weights of the edges as learnable parameters , forming a spatial weight matrix :

[0030] ;

[0031] Based on the spatial weight matrix Calculate the Moran index and build the spatial graph adjacency matrix , as shown in formulas (3) and (4):

[0032] (3)

[0033] (4)

[0034] and Indicates Dimensional Geographical Elements In the geographical unit and The value on for Geographical unit Dimensional Geographical Elements The average value of . The value comes from the spatial weight matrix .

[0035] Preferably, in step 3, based on Geographical units in Geographical elements and wind speed, forecast Geographical units in wind speed;

[0036] The wind speed prediction graph neural network model includes: input layer, two temporal convolution layers, two fully connected layers, graph convolution layer, and graph attention layer; for , is the number of samples used to update the model weights in one training; is the feature dimension, which refers to the number of features in the input data; is the length of the time series; is the number of nodes, and the number of nodes is taken as the number of geographic units , will not change due to the action of the neural network.

[0037] The input layer is the first layer of the neural network, which receives raw input data; in the case of geographic unit wind speed prediction, the input data includes geographic features and wind speed. The input layer does not perform any calculations, but simply passes the data to the next layer;

[0038] Temporal convolutional layers are specifically used to process time series data. They extract temporal correlations through convolution operations. Multiple temporal convolutional layers are stacked together to capture features at different time scales.

[0039] The fully connected layer is used to map the output of the previous layer to a high-dimensional space for more complex nonlinear transformations. After the temporal convolutional layer extracts temporal features, the fully connected layer further extracts and combines these features to provide richer information for the final prediction task.

[0040] The graph convolution layer is used to process graph structure data. In the wind speed prediction of geographic units, the graph convolution layer is used to capture the correlation between different geographic units. This layer updates the representation of nodes by aggregating and transforming the neighbor information of nodes, thereby learning the mutual influence between geographic units.

[0041] Graph attention layer is another form of graph neural network, which allows the model to learn the importance or weights between different nodes in the graph. Through the attention mechanism, the model can identify the most important nodes and relationships for the prediction task, thereby improving the accuracy of the prediction;

[0042] Among them, the attribute graph adjacency matrix Spatial Graph Adjacency Matrix Input to the graph convolution layer and graph attention layer respectively;

[0043] The calculation formula of graph convolution in the graph convolution layer is shown in formulas (5)-(7):

[0044] (5)

[0045] (6)

[0046] (7)

[0047] in, Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Indicates The parameters of the layer convolution, represents the activation function. And the matrix The component is a Laplacian matrix. In is the identity matrix, that is, a matrix with 1s on the diagonal and 0s elsewhere. is a degree matrix, only the main diagonal elements are not zero; express Main diagonal elements; express Each element in; i represents the row number, j represents the column number.

[0048] The calculation formula of the graph attention in the graph attention layer is shown in formulas (8)-(10), which is used to calculate the node Its neighboring nodes ,satisfy , take the attention coefficient and weighted sum as an example.

[0049] (8)

[0050] (9)

[0051] (10)

[0052] in, Representation Node In the The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Represents the activation function. Indicates The parameters of the layer convolution, Indicates the use of shared parameters right Perform a linear transformation, Represents the node and , , the transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. is the activation function.

[0053] The present invention further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and the processor implements the steps in the above method when executing the program.

[0054] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method are implemented.

[0055] The present invention also proposes a method for measuring the attribute similarity between geographical units, and uses a graph convolution module to aggregate the geographical element features and wind speed features of geographical units with attribute similarity in a past time sequence to achieve feature extraction, aiming to enable the model to capture the impact of the wind speed conditions that have appeared in the past time sequence on the future wind speed of geographical units with attribute similarity.

[0056] A distance threshold is set to measure whether there is spatial correlation between geographic units. Edges are established between geographic unit nodes whose Euclidean spatial distance is less than the set threshold, and the weights of the edges are set as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for the subsequent graph attention module, aiming to effectively aggregate the geographic elements and wind speed characteristics of adjacent geographic units according to the weights.

[0057] By introducing an attribute graph that represents the attribute similarity between geographic units and a spatial graph of spatial correlation, and combining two deep learning technologies that focus on graph structure, graph convolutional neural network (GCN) and graph attention neural network (GAT), we fully utilize the multi-source geographic element characteristics and historical wind speed characteristics of geographic units with similar attributes and spatial correlation, so as to better understand the impact of attribute relationships and spatial relationships between geographic units on the model's prediction of future wind speeds, thereby improving the accuracy of wind speed prediction.

[0058] The present invention has the following beneficial technical effects:

[0059] The method proposed in the present invention establishes an attribute graph representing the attribute similarity between geographic units and a spatial graph of spatial correlation, and combines two deep learning technologies focusing on graph structure, namely graph convolutional neural network (GCN) and graph attention neural network (GAT), to comprehensively consider the influence of multi-source geographic elements of geographic units with similar attributes and spatial correlation on the wind speed of the target geographic unit. This forms a new method that is different from only considering the geographic elements of the predicted target geographic unit and using shared weights while ignoring spatial heterogeneity for spatial convolution, thereby better understanding the wind speed law of the geographic unit and providing a technical approach to achieve more accurate wind speed prediction.

[0060] The present invention innovatively applies the third law of geography to the task of predicting wind speed in geographic units. According to the third law of geography, the present invention believes that when the model predicts the wind speed of a geographic unit, it will be affected by the known wind speed fluctuations of geographic units with similar attributes, while the prior art does not model this influence. The present invention uses a graph convolution module to aggregate the geographic element features and wind speed features of geographic units with similar attributes in a past time sequence to achieve feature extraction, aiming to enable the model to capture the impact of the wind speed conditions that have appeared in the past time sequence on the future wind speed of geographic units with similar attributes.

[0061] The prior art uses shared weights for spatial convolution, but ignores the important impact of spatial heterogeneity between regions on the prediction of target wind speed. Unlike the graph model used in the prior art, the present invention sets a distance threshold to measure whether there is spatial correlation between geographic units. Edges are established between geographic unit nodes whose Euclidean space distance is less than the set threshold, and the weights of the edges are set as learnable parameters to form a spatial graph. The present invention uses the spatial graph as an adjacency matrix for subsequent graph attention modules, and can effectively aggregate the geographic elements and wind speed characteristics of adjacent geographic units according to the weights to enhance the ability to predict the wind speed trend of the target geographic unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Target area for wind speed prediction Schematic diagram;

[0063] Figure 2 It is a characteristic diagram of multi-source geographic elements;

[0064] Figure 3 It is a schematic diagram of the method of obtaining values ​​of multi-source geographic elements of geographic units;

[0065] Figure 4 To divide the geographical unit The dimensional feature map is a node Dimensional feature diagram;

[0066] Figure 5 is the attribute graph adjacency matrix Schematic diagram of the construction method;

[0067] Figure 6 is the spatial graph adjacency matrix Schematic diagram of the construction method;

[0068] Figure 7 Schematic diagram of the neural network model structure for wind speed prediction. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above-mentioned purpose, the present invention adopts the following technical scheme.

[0070] The present invention proposes a method for predicting low wind speed events of a wind turbine based on a graph convolutional neural network, which specifically includes the following steps:

[0071] Step 1: constructing a geographic unit attribute map based on the similarity of multi-source geographic elements;

[0072] The present invention takes a single geographic unit as a node in the graph model, and takes multi-source geographic elements that affect the occurrence of low wind speed events, including air pressure (Pa), ambient temperature (°C), dew point temperature (°C), latitudinal wind speed at a height of 10 meters (m / s), longitudinal wind speed at a height of 10 meters (m / s), slope, aspect, elevation and surface cover type, a total of 9 dimensional geographic elements as attribute information of the geographic unit.

[0073] The wind speed forecast target area Divide in space into square geographic units, the resolution of the geographic unit is (m 2 ).like Figure 1 shown.

[0074] Each geographic unit has a characteristic feature that describes the multi-source geographic elements that affect the occurrence of low wind speed events. Dimension. Figure 2 shown.

[0075] The geographical unit determines the value of the multi-source geographical factors according to the size of the area that intersects with the spatial range of the multi-source geographical factors.

[0076] By plot composition; set The areas are , the temperature values ​​are ,like Figure 3 shown.

[0077] Geographical unit The area is , the temperature value is , the calculation method is shown in formula (1):

[0078] (1)

[0079] Map geographic units to nodes. The dimensional feature map is a node dimensional features. Figure 4 shown.

[0080] There are four research target geographical units in the region indivual( ), to measure the geographical units With geographical units The present invention measures the similarity of attributes between each dimension of geographical elements in a certain time series. The similarity of, and sum, and then find the average value, as a geographical unit With geographical units The similarity of attributes between , as shown in formula (2):

[0081] (2)

[0082] In the formula, Represents the dimension of geographic features, To record geographical units exist Dimensions a vector of sequence values, is the covariance between two vectors, is the variance of a vector.

[0083] Calculate attribute similarity between geographic units separately As the weight of the edge between nodes, build the adjacency matrix of the attribute graph ,like Figure 5 shown.

[0084] .

[0085] Step 2, constructing a spatial map of geographic units based on Euclidean spatial distance;

[0086] The present invention uses a single geographic unit as a node in the graph model. The present invention sets a distance threshold To measure whether there is spatial correlation between geographic units. If the Euclidean distance between two geographic units Greater than the set threshold , then when predicting the wind speed trend of one geographical unit, the spatial correlation between the other geographical unit and it is not considered.

[0087] Distance in Euclidean space Less than the set threshold Establish edges between geographic unit nodes and set the weights of the edges as learnable parameters , forming a spatial weight matrix .

[0088] ;

[0089] Based on the spatial weight matrix Calculate the Moran index and build the spatial graph adjacency matrix , as shown in formulas (3), (4) and Figure 6 shown.

[0090] (3)

[0091] (4)

[0092] and Indicates Dimensional Geographical Elements In the geographical unit and The value on . for Geographical unit Dimensional Geographical Elements The average value of . The value comes from the spatial weight matrix .

[0093] Step 3: Neural network modeling of wind speed prediction graph;

[0094] based on Geographical units in (Length is time series) of geographical elements and wind speed, predict Geographical units in (Length is time series) of wind speed.

[0095] The structure of the wind speed prediction graph neural network model is as follows Figure 7 As shown in the figure, the wind speed prediction graph neural network model includes: input layer, two time convolution layers, two fully connected layers, graph convolution layer, and graph attention layer. for , is the number of samples used to update the model weights in one training. is the feature dimension, which refers to the number of features in the input data. In the present invention, the initial value is 13, which means that each sample has 13-dimensional geographic elements as input features. The feature dimension determines the size of the input layer and is the basis for the amount of information that the model can capture. is the length of the time series. In time series forecasting, longer time series can provide more time-dependent information, but will increase computational complexity. The choice of time series length depends on the requirements of the specific task and the amount of available data. is the number of nodes, in this invention the number of nodes is the number of geographical units , will not change due to the action of the neural network.

[0096] Input layer: The input layer is the first layer of the neural network and is responsible for receiving raw input data. In the case of geographic unit wind speed forecasting, the input data includes geographic features and wind speed. The input layer does not perform any calculations, but simply passes the data to the next layer.

[0097] Temporal convolutional layers: Temporal convolutional layers are specifically designed to process time series data. They extract temporal correlations through convolution operations. There are usually multiple temporal convolutional layers stacked together to capture features at different time scales.

[0098] Fully connected layer: maps the output of the previous layer to a high-dimensional space for more complex nonlinear transformations. After the temporal convolutional layer extracts temporal features, the fully connected layer can further extract and combine these features to provide richer information for the final prediction task.

[0099] Graph convolution layer: The graph convolution layer is used to process graph structured data. In the wind speed prediction of geographic units, the graph convolution layer can capture the correlation between different geographic units. This layer updates the representation of nodes by aggregating and transforming the neighbor information of nodes, thereby learning the mutual influence between geographic units.

[0100] Graph Attention Layer: Graph Attention Layer is another form of Graph Neural Network that allows the model to learn the importance or weights between different nodes in the graph. Through the attention mechanism, the model can identify the most important nodes and relationships for the prediction task, thereby improving the accuracy of the prediction.

[0101] Among them, the attribute graph adjacency matrix Spatial Graph Adjacency Matrix Input to the graph convolution layer and graph attention layer respectively.

[0102] The calculation formula of graph convolution in the graph convolution layer is shown in formulas (5)-(7):

[0103] (5)

[0104] (6)

[0105] (7)

[0106] in, Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Indicates The parameters of the layer convolution, represents the activation function. And the matrix The component is a Laplacian matrix. In is the identity matrix, that is, a matrix with 1s on the diagonal and 0s elsewhere. is a degree matrix, only the main diagonal elements are not zero; express Main diagonal elements; express Each element in; i represents the row number and j represents the column number.

[0107] The calculation formula of the graph attention in the graph attention layer is shown in formulas (8)-(10), which is used to calculate the node Its neighboring nodes (satisfy ) and weighted sum of the attention coefficients.

[0108] (8)

[0109] (9)

[0110] (10)

[0111] in, Representation Node In the The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Represents the activation function. Indicates The parameters of the layer convolution, Indicates the use of shared parameters right Perform a linear transformation, Represents the node and ( ) The transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. It is a kind of activation function.

[0112] Step 4: Sample data set construction and model optimization; collection of research objectives Geographical units in (Length is The dataset consists of geographical elements and historical wind speed data (time series). Model input: (1) Geographical units in (Length is (2) Based on the geographical elements of geographical units in a certain time series The similarity is established ; (3) Based on the Euclidean spatial distance of geographical units ; Model tag: (1) Geographical units in (Length is time series) of wind speed.

[0113] Build the sample as above The samples are divided into training set, validation set and test set in a 6:2:2 ratio. The training set is used to train the model; the validation set is used to adjust the model hyperparameters; and the test set is used to test the final accuracy of the model.

[0114] When applying the model, collect data and build Geographical units in (Length is The feature matrix, attribute graph and spatial graph of the time series are input into the trained model for forward propagation to obtain Geographical units in (Length is The wind speed prediction results of the time series.

[0115] According to the task scenario requirements, the length is In the time series, the wind speed is continuously lower than The geographical unit with a wind speed of less than 100 km / h (m / s) is judged as a low wind speed event.

Claims

1. A method for predicting low wind speed events of wind turbines based on graph convolutional neural networks. It is characterized in that The method specifically comprises the following steps: Step 1: construct a geographic unit attribute graph based on the similarity of multi-source geographic elements; use a single geographic unit as a node in the graph model, and use the multi-source geographic elements that affect the occurrence of low wind speed events as the attribute information of the geographic unit; measure the attribute similarity between geographic units in a certain time series as the weight of the edge between nodes, the higher the similarity, the higher the weight, thereby completing the construction of the attribute graph; Step 2: construct a spatial graph of geographic units based on Euclidean spatial distance; use a single geographic unit as a node in the graph model, set a distance threshold to measure whether there is spatial correlation between geographic units, and if the Euclidean spatial distance between two geographic units is greater than the set threshold, then when predicting the wind speed trend of one of the geographic units, the spatial correlation between the other geographic unit and it is not considered, and establish edges between the geographic unit nodes whose Euclidean spatial distance is less than the set threshold, and set the weight of the edge as a learnable parameter to form a spatial graph; Step 3: Modeling the wind speed prediction graph neural network; using the attribute graph as the adjacency matrix for the graph convolution layer, using the spatial graph as the adjacency matrix for the graph attention layer, and convolving the time dimension to achieve end-to-end wind speed time series prediction; Step 4: Sample data set construction and model optimization: construct a time series sample data set of geographic units including air pressure, ambient temperature, dew point temperature, zonal wind speed at a height of 10 meters, and meridional wind speed at a height of 10 meters for model training, verification, and testing.

2. The method according to claim 1, It is characterized in that In step 1, the multi-source geographical elements that affect the occurrence of low wind speed events, including air pressure, ambient temperature, dew point temperature, latitudinal wind speed at a height of 10 meters, longitudinal wind speed at a height of 10 meters, slope, aspect, elevation and land cover type, a total of 9 dimensional geographical elements are used as attribute information of geographical units; the wind speed prediction target area Divide in space into square geographic units, the resolution of the geographic unit is m 2 ; By plot composition; set The areas are , the temperature values ​​are , set geographic unit The area is , the temperature value is , the calculation method is shown in formula (1): (1) The geographical unit The dimensional feature map is a node dimensional features, assuming that there are a total of geographical units in the region indivual, , to measure the geographical units With geographical units Taking the attribute similarity between two dimensions as an example, we measure the similarity of each dimension of geographical elements in a certain time series. The similarity of the two is summed up and averaged as the geographical unit. With geographical units The similarity of attributes between , as shown in formula (2): (2) In the formula, Represents the dimension of geographic features, To record geographical units exist Dimensions a vector of sequence values, , Record geographical units With geographical units exist Dimensions a vector of sequence values, is the covariance between two vectors, is the variance of a vector; Calculate attribute similarity between geographic units separately As the weight of the edge between nodes, build the adjacency matrix of the attribute graph : 。 3. The method according to claim 1, It is characterized in that In step 2, set the distance threshold To measure whether there is spatial correlation between geographic units, if the Euclidean distance between two geographic units Greater than the set threshold , then when predicting the wind speed trend of one geographical unit, the spatial correlation between the other geographical unit and it is not considered; in the Euclidean space distance Less than the set threshold Establish edges between geographic unit nodes and set the weights of the edges as learnable parameters , forming a spatial weight matrix : ； Based on the spatial weight matrix Calculate the Moran index and build the spatial graph adjacency matrix , as shown in formulas (3) and (4): (3) (4) and Indicates Dimensional Geographical Elements In the geographical unit and The value on for Geographical unit Dimensional Geographical Elements The average value of The value comes from the spatial weight matrix .

4. The method according to claim 1, It is characterized in that In step 3, based on Geographical units in Geographical elements and wind speed, forecast Geographical units in wind speed; The wind speed prediction graph neural network model includes: input layer, two temporal convolution layers, two fully connected layers, graph convolution layer, and graph attention layer; for , is the number of samples used to update the model weights in one training; is the feature dimension, which refers to the number of features in the input data; is the length of the time series; is the number of nodes, and the number of nodes is taken as the number of geographic units , will not change due to the action of the neural network; The input layer is the first layer of the neural network, which is used to receive raw input data. In the case of wind speed prediction in geographic units, the input data includes geographic elements and wind speed. The input layer does not perform any calculations, but simply passes the data to the next layer. Temporal convolutional layers are specifically used to process time series data. They extract temporal correlations through convolution operations. Multiple temporal convolutional layers are stacked together to capture features at different time scales. The fully connected layer is used to map the output of the previous layer to a high-dimensional space for more complex nonlinear transformations. After the temporal convolutional layer extracts temporal features, the fully connected layer further extracts and combines these features to provide richer information for the final prediction task. The graph convolution layer is used to process graph structure data. In the wind speed prediction of geographic units, the graph convolution layer is used to capture the correlation between different geographic units. This layer updates the representation of nodes by aggregating and transforming the neighbor information of nodes, thereby learning the mutual influence between geographic units. Graph attention layer is another form of graph neural network, which allows the model to learn the importance or weights between different nodes in the graph. Through the attention mechanism, the model can identify the most important nodes and relationships for the prediction task, thereby improving the accuracy of the prediction; Among them, the attribute graph adjacency matrix Spatial Graph Adjacency Matrix Input to the graph convolution layer and graph attention layer respectively; The calculation formula of graph convolution in the graph convolution layer is shown in formulas (5)-(7): (5) (6) (7) in, Indicates that the node is The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, Indicates The parameters of the layer convolution, represents the activation function, and the matrix The component is a Laplace moment, In is the identity matrix, that is, a matrix with 1 on the diagonal and 0 on the rest. is a degree matrix, only the main diagonal elements are not zero; express Main diagonal elements; express Each element in; i represents the row number, j represents the column number; The calculation formula of the graph attention in the graph attention layer is shown in formulas (8)-(10), which is used to calculate the node Its neighboring nodes ,satisfy , take the attention coefficient and weighted sum as an example: (8) (9) (10) in, Representation Node In the The feature vector of the layer, It means that after convolution, the node is The feature vector of the layer, represents the activation function, Indicates The parameters of the layer convolution, Indicates the use of shared parameters right Perform a linear transformation, Represents the node and , , the transformed results are spliced, It means mapping the concatenated high-dimensional features to a real number. is the activation function.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, It is characterized in that When the processor executes the program, the steps in the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 4 are implemented.

Images