A method for predicting the wind speed of wind turbines in a wind farm

Abstract

The present invention discloses a method for predicting the wind speed of wind farm units. Aiming at the problems of low prediction accuracy, inapplicability to big data processing, and complex modeling process existing in the existing wind farm unit wind speed prediction technology; the present invention comprises the following steps: Step 1, obtaining the information of each unit in the wind farm and preprocessing the data; Step 2, spatial feature extraction based on a graph network; Step 3, temporal feature extraction based on a gated recurrent unit neural network; Step 4, spatio-temporal feature fusion to realize the prediction of the wind speed of each wind turbine in the wind farm. The advantage is that a technical solution is constructed by considering the spatial distribution characteristics of the wind farm units and the temporal characteristics of the wind speed data in combination, and a method for predicting the wind speed of wind farm units is proposed, realizing the effective prediction of the wind speed of all the units in the whole field.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power, and particularly to a method for predicting the wind speed of wind turbines in a wind farm. Background Art

[0002] Wind speed, as an important measurement index of wind energy, has an important impact on the control of wind turbines and the power generation of wind farms. Therefore, predicting the wind speed of wind farm turbines and improving the prediction accuracy of wind turbine wind speed are of great significance for the control optimization of wind turbines and power grid dispatching.

[0003] The existing methods for predicting the wind speed of wind farm turbines can be mainly divided into the following categories: 1. Wind speed prediction technologies based on mathematical statistics analysis, such as autoregressive models, ARIMA models, etc. These technologies mainly achieve the prediction of wind speed data based on the analysis of the relationship of historical time series data; 2. Wind speed prediction technologies based on machine learning models, such as support vector machines, backpropagation neural networks, Xgboost, etc. These technologies mainly achieve the prediction of wind speed data through data feature mining; 3. Wind speed prediction technologies based on mechanism models, which mainly achieve the prediction of wind speed by modeling the influencing factors of wind speed.

[0004] Although the above methods can all achieve the prediction of the wind speed of wind farm turbines, they also have certain limitations. For example, the wind speed prediction technology based on mathematical statistics analysis is easily affected by abnormal data in historical data and is not applicable to big data processing; the wind speed prediction technology based on machine learning models has stronger data feature analysis ability than mathematical statistics technology, but most technologies still only consider the influence of time features and cannot mine the deep features of data; the wind speed prediction technology based on mechanism models requires a large amount of data collection and the calculation process is complex.

[0005] For example, a "Method for Predicting Wind Speeds at Multiple Points in a Wind Farm Based on a Convolutional Recurrent Neural Network" disclosed in a Chinese patent document, with the publication number CN110889535A and the publication date of March 17, 2020, includes the following steps: Step 1: Collect the operation data of the wind farm, where the collected data includes the measured wind speeds and measured wind directions at the positions of multiple wind turbines; Step 2: Establish a convolutional module of the wind speed prediction model for multiple points in the wind farm based on the convolutional recurrent neural network according to the data collected in Step 1; Step 3: Establish an LSTM module of the wind speed prediction model for multiple points in the wind farm based on the convolutional recurrent neural network according to Step 1; Step 4: Connect the outputs of the convolutional module and the LSTM module; Step 5: Train the neural network model with the mean absolute error (MAE) loss function index. This technical solution uses the LSTM module of the wind speed prediction model for multiple points in the wind farm based on the convolutional recurrent neural network. The LSTM module has disadvantages in parallel processing, and when dealing with too much data, it will lead to problems of low prediction accuracy and time-consuming calculation. Summary of the Invention

[0006] To overcome the problems in the existing wind farm unit wind speed prediction technology, such as low prediction accuracy, inapplicability to big data processing, and complex modeling process, from the perspective of data-driven, a technical solution is constructed by combining the potential spatial characteristics of wind farm units and the time characteristics of wind speed data, and a method for predicting the wind speed of wind farm units is provided.

[0007] The above technical problems of the present invention are mainly solved by the following technical solutions:

[0008] The present invention includes the following steps:

[0009] Step 1: Obtain the information of each unit in the wind farm and preprocess the data: Collect the geographical location information and wind speed information of each unit in the wind farm, and preprocess the geographical location information and wind speed information.

[0010] Step 2: Spatial feature extraction based on graph network: Construct a graph network based on the geographical location information of each unit in the wind farm, and extract the spatial feature information of the wind farm units through graph convolution operations.

[0011] Step 3: Temporal feature extraction based on gated recurrent unit neural network: Based on the wind speed information of each unit in the wind farm, extract the temporal feature information of the wind speed information of the wind farm units through the gated recurrent unit neural network.

[0012] Step 4: Spatiotemporal feature fusion to realize the prediction of the wind speed of each wind turbine in the wind farm: Perform feature fusion processing on the spatial features of each unit in the wind farm and the temporal features of the wind speed of each unit, and realize the prediction of the wind speed of each wind turbine in the wind farm through a multi-layer fully connected layer network based on the spatiotemporal fusion features.

[0013] This solution constructs a technical solution by considering both the influence of the spatial distribution of wind speeds in a wind farm and the trend of wind speed changes. It uses the location information of the wind turbines in the wind farm to construct a wind turbine network in the wind farm, extracts the spatial correlations between turbines, and obtains the characteristics of the spatial distribution of wind speeds in the wind farm. It uses the wind speed information of the wind turbines in the wind farm, based on a gated recurrent unit neural network, to extract the temporal characteristics of the wind speed sequences of each turbine and obtains the characteristics of the wind speed change trends of each turbine in the wind farm. Finally, it combines the spatial characteristics and temporal characteristics to achieve an effective prediction of the wind speeds of each turbine in the wind farm.

[0014] Preferably, in step 1, the geographical location information of each wind turbine in the wind farm is obtained, and the data is preprocessed: the longitude and latitude of the geographical location information of each wind turbine in the wind farm are obtained, and the geographical location information is preprocessed.

[0015] It is assumed that the number of wind turbines in the target wind farm is N, and the set of geographical location information of each turbine is P = {Lat i , Lon i}. The set of wind speed information of each turbine is V = {V i}. Among them, 1 ≤ i ≤ N, Lat i represents the latitude of the i-th turbine, Lon i represents the longitude of the i-th turbine, and V i represents the set of wind speed information of the i-th turbine;

[0016] Based on the above settings, distance calculation processing is performed on the geographical location information of the turbines to construct a distance set Dis = {d ij}, 1 ≤ i ≤ N, 1 ≤ j ≤ N, and the calculation formula for d ij is as follows:

[0017]

[0018] ΔLat′ ij - Lat′ i - Lat′ j

[0019] ΔLon′ ij = Lon′ i - Lont′ j

[0020]

[0021] In the formula, temp ij is an intermediate calculation variable, and d ij is the distance between wind turbine i and wind turbine j.

[0022] This solution can calculate the geographical location information of the turbines and preprocess the data information, thus making the data more standardized and facilitating subsequent calculations.

[0023] Preferably, in step 1, the wind speed information of each unit in the wind farm is obtained, and the data is preprocessed: the wind speed information of the units in the wind farm is obtained, and the wind speed data of the units is subjected to standardization processing and serialization processing;

[0024] The standardization calculation formula is as follows:

[0025]

[0026] In the formula, V i ={v it}, where v it is the wind speed data of the i-th unit at time t, and v' it is the standardized wind speed data of the i-th unit at time t;

[0027] The result of serialization processing is as follows:

[0028]

[0029] In the formula, n is the number of wind speed data of each unit, T is the size of the historical time window, P is the size of the prediction window, X is the result after serialization of the wind turbine speed data, and S is the set of predicted wind speed data of the unit.

[0030] Adopting this solution can calculate the wind speed information of each unit in the wind farm, and perform standardization processing and serialization processing on the data information, so as to make the data more standardized.

[0031] Preferably, the specific content in step 2 is as follows:

[0032] Step 2-1: Construct a graph network based on the geographical location information of the wind farm units:

[0033] Taking the wind turbines in the wind farm as nodes V and the straight-line paths of the wind turbines as edges E, a wind turbine graph G=(V, E, A) is constructed, where A is the adjacency matrix of graph G;

[0034] According to the calculation result Dis of the distances between the wind turbines in the wind farm, construct the adjacency matrix A of the wind turbines. The construction process is as follows:

[0035]

[0036] In the formula, σ is the variance of Dis;

[0037] Step 2-2: Extract the spatial feature information of the wind farm units based on the graph convolutional network:

[0038] Set the input data of the graph convolutional network as x, then the calculation process of the graph convolutional network is as follows:

[0039] g θ*G x = g θ(L)x = g θ (U ∧ U T )x = Ug θ (∧)U T x

[0040] L = D - A

[0041]

[0042] where g θ is the Laplace function with parameter θ, * G is the graph convolution operation, L is the Laplacian matrix, A is the adjacency matrix, D is the degree matrix, U is an orthogonal matrix, and U T is the transpose of the orthogonal matrix, and ∧ is the diagonal matrix composed of the eigenvalues of L.

[0043] This solution is adopted to obtain the spatial feature information of the wind farm units based on the graph network.

[0044] Preferably, when the graph network is large and the number of wind farm unit data is large, the solution time of the eigenvalues of the Laplacian matrix is relatively long. Therefore, Chebyshev polynomials are used to solve the above problems, and the calculation process is as follows:

[0045]

[0046] where T k (x) = 2xT k-1 (x) - T k-2 (x), T0(x) = 1, T1(x) = x, K = 2 is the order of the Chebyshev polynomial, θ k is the coefficient in front of the k-th polynomial in the Chebyshev polynomial, L′ is the normalized Laplacian matrix, λ max is the largest eigenvalue of the Laplacian matrix L, and I is the identity matrix.

[0047] This solution is adopted because when the graph network is large, that is, when the number of wind farm unit data is large, the corresponding solution process will be too complex. Therefore, Chebyshev polynomials are used to simplify the solution process.

[0048] Preferably, the specific content in step 3 is as follows:

[0049] Step 3-1: Construct a gated recurrent unit neural network:

[0050] Assume the input of the gated recurrent unit neural network is x t , then the calculation process of the gated recurrent unit neural network is as follows:

[0051] z t = σ′(W z x t + Uz h t-1 )

[0052] r t = σ'(W r x t + U r h t-1 )

[0053]

[0054] In the formula, σ' is the sigmoid activation function, o is the dot product calculation, z t is the output result of the update gate, r t is the output result of the reset gate, h t is the output result of the hidden layer, W z , W r , W h are the weight matrices of the update gate, reset gate, and cell unit under the input of x t respectively, and U z , U r , U h are the weight matrices of the update gate, reset gate, and cell unit under the input of h t-1 respectively;

[0055] Step 3-2: Extract the time features of the wind turbine speed data based on the gated recurrent unit neural network:

[0056] After serializing the wind turbine speed data, the result X is divided into N groups according to the units, that is, X = {X1, X2,..., X N}), and the time features are extracted from the wind speed data set X i of each unit in the wind farm respectively. The extraction process is as follows:

[0057] H i = GRU(X i )

[0058] H = {Hi}

[0059] In the formula, H i is the time feature extracted from the i-th unit, GRU represents the gated recurrent unit neural network, and H is the set of time features extracted from the wind farm units.

[0060] This solution is adopted to extract the time features of the wind speed information of the wind farm units through the gated recurrent unit neural network based on the wind speed information of each unit in the wind farm.

[0061] Preferably, the specific content in step 4 is:

[0062] Step 4-1: Feature fusion processing:

[0063] The time characteristics and spatial characteristics of wind speed of wind farm units are fused. The fusion process is as follows:

[0064] Z=H+g θ*G X

[0065] Step 4-2: Output of wind speed prediction for each unit in the wind farm:

[0066] Based on the fusion features and the multi-layer fully connected layer output prediction results, the calculation process is as follows:

[0067] Y=W y2 (W y1 Z+b y1 )+b y2

[0068] Where W y1 , b y1 is the weight and bias of the first fully connected layer, W y2 , b y2 The weights and biases of the second fully connected layer, Y is the predicted result output.

[0069] This scheme is adopted to fuse the spatial characteristics of each unit in the wind farm with the temporal characteristics of the wind speed of each unit, and to predict the wind speed of each wind turbine in the wind farm through a multi-layer fully connected layer network based on the spatiotemporal fusion characteristics.

[0070] The beneficial effects of the present invention are: 1. The present invention constructs a technical solution by considering the spatial distribution characteristics of wind farm units and the temporal characteristics of wind speed data, and proposes a wind speed prediction method for wind farm units, thereby realizing effective prediction of the wind speed of all units in the field; 2. In the present invention, a wind farm wind turbine graph network is constructed with wind turbines as nodes to mine the spatial characteristics of wind speed of wind farm units. Compared with a single method based on statistical analysis of temporal characteristics, the mining and fusion of spatial characteristics will further improve the accuracy of wind speed prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of the present invention.

[0072] Figure 2 It is a framework diagram of the wind speed prediction model of the wind farm unit of the present invention. DETAILED DESCRIPTION

[0073] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0074] Example:

[0075] A method for predicting wind speed of a wind turbine generator set in a wind farm according to this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0076] Step 1. Obtain the information of each unit in the wind farm and preprocess the data: Collect the geographical location information and wind speed information of each unit in the wind farm, and preprocess the geographical location information and wind speed information;

[0077] Step 1-1. Obtain the geographical location information of each unit in the wind farm and preprocess the data: Obtain the longitude and latitude of the geographical location information of each unit in the wind farm, and preprocess the geographical location information;

[0078] Set the number of wind turbines in the target wind farm to N, the set of geographical location information of each unit to P = {Lat i , Lon i}, and the set of wind speed information of each unit to V = {V i}. Among them, 1 ≤ i ≤ N, Lat i represents the latitude of the i-th unit, Lon i represents the longitude of the i-th unit, and V i represents the set of wind speed information of the i-th unit;

[0079] Based on the above settings, perform distance calculation processing on the geographical location information of the units, and construct a distance set Dis = {d ij}, 1 ≤ i ≤ N, 1 ≤ j ≤ N, and the calculation formula of d ij is as follows:

[0080]

[0081] ΔLat′ ij = Lat′ i - Lat′ j

[0082] ΔLon′ ij = Lon′ i - Lont′ j

[0083]

[0084] In the formula, temp ij is an intermediate calculation variable, and d ij is the distance between wind turbine i and wind turbine j;

[0085] Step 1-2. Obtain the wind speed information of each unit in the wind farm and preprocess the data: Obtain the wind speed information of the units in the wind farm, and perform standardization processing and serialization processing on the wind speed data of the units;

[0086] The standardization calculation formula is as follows:

[0087]

[0088] wherein, V i ={v it}, where v it is the wind speed data of the i-th unit at time t, and v' it is the normalized wind speed data of the i-th unit at time t;

[0089] The serialization processing results are as follows:

[0090]

[0091] wherein, n is the number of wind speed data of each unit, T is the size of the historical time window, P is the size of the prediction window, X is the result after serialization of the wind speed data of the wind turbine, and S is the set of predicted wind speed data of the unit.

[0092] Step 2: Spatial feature extraction based on the graph network: Construct a graph network based on the geographical location information of each unit in the wind farm, and extract the spatial feature information of the units in the wind farm through graph convolution operations; the specific content is as follows:

[0093] Step 2-1: Construct a graph network based on the geographical location information of the units in the wind farm:

[0094] Taking the wind turbines in the wind farm as nodes V and the direct paths of the wind turbines as edges E, construct a wind turbine graph G=(V, E, A), where A is the adjacency matrix of graph G;

[0095] Construct the adjacency matrix A of the wind turbines according to the calculation result Dis of the distances between the wind turbines in the wind farm, and the construction process is as follows:

[0096]

[0097] wherein, σ is the variance of Dis;

[0098] Step 2-2: Extract the spatial feature information of the units in the wind farm based on the graph convolutional network:

[0099] Set the input data of the graph convolutional network as x, then the calculation process of the graph convolutional network is as follows:

[0100] g θ*G x = g θ (L)x = g θ (U∧U T )x = Ug θ (∧)U T x

[0101] L = D - A

[0102]

[0103] wherein, g θis the Laplace function with parameter θ, * G is the graph convolution operation, L is the Laplace matrix, A is the adjacency matrix, D is the degree matrix, U is the orthogonal matrix, U T is the transpose of the orthogonal matrix, ∧ is the diagonal matrix composed of the eigenvalues of L;

[0104] Considering that when the graph network is large and there is a large amount of data of wind farm units, the eigenvalue solution time of the corresponding Laplace matrix is long. Therefore, Chebyshev polynomials are used to solve the above problems, and the calculation process is as follows:

[0105]

[0106] In the formula, T k (x) = 2xT k-1 (x) - T k-2 (x), T0(x) = 1, T1(x) = x, K = 2 is the order of the Chebyshev polynomial, θ k is the coefficient in front of the k-th polynomial in the Chebyshev polynomial, L′ is the normalized Laplace matrix, λ max is the largest eigenvalue of the Laplace matrix L, and I is the identity matrix.

[0107] Step 3, Time feature extraction based on the gated recurrent unit neural network: Based on the wind speed information of each unit in the wind farm, the time features of the wind speed information of the wind farm units are extracted through the gated recurrent unit neural network; the specific content is as follows:

[0108] Step 3-1, Construct the gated recurrent unit neural network:

[0109] Assume that the input of the gated recurrent unit neural network is x t , then the calculation process of the gated recurrent unit neural network is as follows:

[0110] z t = σ′(W z x t + U z h t-1 )

[0111] r t = σ′(W r x t + U r h t-1 )

[0112]

[0113] In the formula, σ′ is the sigmoid activation function, o is the dot product calculation, z t is the output result of the update gate, r t is the output result of the reset gate, ht is the output result of the hidden layer, W z , W r , W h are the weight matrices for the update gate, reset gate, and x in the cell unit t input, respectively; U z , U r , U h are the weight matrices for the update gate, reset gate, and h in the cell unit t-1 input, respectively;

[0114] Step 3-2: Extract the time features of the wind speed data of the wind turbine based on the gated recurrent unit neural network:

[0115] After serializing the wind speed data of the wind turbine, the result X is divided into N groups according to the units, i.e., X = {X1, X2,..., X N}, and the time features are extracted from the wind speed data set X of each unit in the wind farm respectively. The extraction process is as follows: i The extraction process is as follows:

[0116] H i = GRU(X i )

[0117] H = {H i}

[0118] In the formula, H i is the time feature extracted from the i-th unit, GRU represents the gated recurrent unit neural network, and H is the set of time features extracted from the wind farm units.

[0119] Step 4: Spatiotemporal feature fusion to realize the prediction of the wind speed of each wind turbine in the wind farm: The spatial features of each unit in the wind farm and the time features of the wind speed of each unit are subjected to feature fusion processing, and the prediction of the wind speed of each wind turbine in the wind farm is realized through a multi-layer fully connected layer network based on the spatiotemporal fusion features; as Figure 2 shown, the specific content is as follows:

[0120] Step 4-1: Feature fusion processing:

[0121] The time features and spatial features of the wind speed of the wind farm units are subjected to feature fusion. The fusion process is as follows:

[0122] Z = H + g θ*G X

[0123] Step 4-2: Prediction output of the wind speed of each unit in the wind farm:

[0124] Based on the fusion features and the output of the multi-layer fully connected layer, the prediction result is calculated as follows:

[0125] Y = W y2 (W y1Z + b y1 ) + b y2

[0126] In the formula, W y1 , b y1 are the weights and biases of the first fully connected layer, and W y2 , b y2 are the weights and biases of the second fully connected layer. Y is the predicted result output.

[0127] In this embodiment, taking the wind turbine as a point, a graph network of wind farm wind turbines is constructed based on the longitude and latitude information of the wind turbine generator set. During the construction of the graph network of wind farm wind turbines, an adjacency matrix calculation method is proposed to replace the traditional {0, 1} adjacency matrix representation method, which better expresses the spatial relationship of the wind turbine generator sets. The wind speed data features of the wind farm wind turbines are extracted from both spatial and temporal aspects. The spatial features are extracted from the graph network of wind farm wind turbines based on graph convolution operations. Considering the extraction of temporal features and spatio-temporal features from the wind speed data of the wind farm wind turbines based on the gated recurrent unit neural network further improves the prediction accuracy of the wind speed of the entire wind farm wind turbines.

[0128] It should be understood that the embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

Claims

1. A method for predicting the wind speed of a wind turbine in a wind farm, characterized in that, It includes the following steps: Step 1: Obtain the information of each unit in the wind farm and preprocess the data. Collect the geographical location information and wind speed information of each unit in the wind farm, and preprocess the geographical location information and wind speed information. Step 2: Spatial feature extraction based on graph network. Construct a graph network based on the geographical location information of each unit in the wind farm, and extract the spatial feature information of the wind farm units through graph convolution operations. Step 3: Temporal feature extraction based on gated recurrent unit neural network. Based on the wind speed information of each unit in the wind farm, extract the temporal feature of the wind speed information of the wind farm units through the gated recurrent unit neural network. Step 4: Spatiotemporal feature fusion to achieve the prediction of the wind speed of each fan in the wind farm. Perform feature fusion processing on the spatial features of each unit in the wind farm and the temporal features of the wind speed of each unit, and based on the spatiotemporal fusion features, realize the prediction of the wind speed of each fan in the wind farm through a multi-layer fully connected layer network.

2. The wind speed prediction method for a wind farm wind turbine generator set according to claim 1, wherein, In step 1, obtain the geographical location information of each unit in the wind farm and preprocess the data: obtain the longitude and latitude of the geographical location information of each unit in the wind farm, and preprocess the geographical location information; set the number of wind turbines in the target wind farm as N, and the set of geographical location information of each unit as P = {Lat i , Lon i}, and the set of wind speed information of each unit as V = {V i}; where 1 ≤ i ≤ N, Lat i represents the latitude of the i-th unit, Lon i represents the longitude of the i-th unit, and V i represents the set of wind speed information of the i-th unit; Based on the above settings, distance calculation processing is performed on the unit's geographical location information to construct a distance set Dis = {d ij} where 1 ≤ i ≤ N, 1 ≤ i ≤ N, d ii The calculation formula is as follows: ΔLat′ ij = Lat′ i - Lat′ j ΔLon′ ij = Lon′ i - Lont′ j where temp ij is an intermediate calculation variable, and d ij is the distance between wind turbine i and wind turbine j.

3. The wind speed prediction method for a wind farm wind turbine generator set according to claim 2, wherein In Step 1, obtain the wind speed information of each unit in the wind farm and preprocess the data: Obtain the wind speed information of the wind farm units, and perform standardization processing and serialization processing on the wind speed data of the units. The standardization calculation formula is as follows: Where, V i = {v it}, where v it is the wind speed data of the i-th unit at time t, and v' it is the normalized wind speed data of the i-th unit at time t; The result of serialization processing is as follows: In the formula, n is the number of wind speed data of each unit, T is the size of the historical time window, P is the size of the prediction window, X is the result after serialization of the wind turbine speed data, and S is the set of predicted wind speed data of the unit.

4. A wind speed prediction method for a wind turbine in a wind farm according to claim 3, wherein The specific content in Step 2 is as follows: Step 2-1: Construct a graph network based on the geographical location information of the wind farm units. Take the wind turbines in the wind farm as nodes V and the straight-line paths of the wind turbines as edges E to construct a wind turbine graph G=(V, E, A), where A is the adjacency matrix of graph G. Construct the adjacency matrix A of the wind turbines according to the calculated result Dis of the distances between the wind turbines in the wind farm. The construction process is as follows: In the formula, σ is the variance of Dis. Step 2-2: Extract the spatial feature information of the wind farm units based on the graph convolutional network. Set the input data of the graph convolutional network as x, then the calculation process of the graph convolutional network is as follows: g e * G x = g e (L)x = g e (U∧U T )x = Ug θ (∧)U T x L = D - A where, g θ is the Laplace function with parameter θ, * G is the graph convolution operation, L is the Laplacian matrix, A is the adjacency matrix, D is the degree matrix, U is an orthogonal matrix, U T is the transpose of the orthogonal matrix, and ∧ is the diagonal matrix composed of the eigenvalues of L.

5. A method for predicting the wind speed of a wind turbine in a wind farm according to claim 4, characterized in that, When the graph network is large and the data of the wind farm units are numerous, the time for solving the eigenvalues of the Laplacian matrix is relatively long. Therefore, the Chebyshev polynomial is used to solve the above problem. The calculation process is as follows: where T k (x) = 2xT k-1 (x) - T k-2 (x), T0(x) = 1, T1(x) = x, K = 2 is the order of the Chebyshev polynomial, θ k is the coefficient in front of the k-th polynomial in the Chebyshev polynomial, L′ is the normalized Laplacian matrix, λ max is the largest eigenvalue of the Laplacian matrix L, and I is the identity matrix.

6. A wind speed prediction method for a wind farm wind turbine generator set according to claim 5, characterized in that, The specific content in Step 3 is as follows: Step 3-1: Construct a gated recurrent unit neural network. Assume that the input of the gated recurrent unit neural network is x t , then the calculation process of the gated recurrent unit neural network is as follows: z t = σ′(W z x t + U z h t-1 ) r t = σ′(W r x t + U r h t-1 ) where, σ′ is the sigmoid activation function, o is the dot product calculation, z t is the output result of the update gate, r t is the output result of the reset gate, h t is the output result of the hidden layer, W z 、W r 、W h are the weight matrices of the update gate, the reset gate, and x t input in the cell unit respectively, U z 、U r 、U h are the weight matrices of the update gate, the reset gate, and h t-1 input in the cell unit respectively; Step 3-2: Extract the temporal features of the wind speed data of the wind turbines based on the gated recurrent unit neural network. After serializing the wind turbine speed data, the result X is divided into N groups according to the turbines, i.e., X = {X1, X2, …, X N}, and the speed data sets X i of each turbine in the wind farm are respectively subjected to time feature extraction, and the extraction process is as follows: H i = GRU(X i ) H = {H i} where H i is the time feature extracted by the i-th unit, GRU represents the gated recurrent unit neural network, and H is the set of time features extracted by the wind farm units.

7. A method for predicting the wind speed of a wind turbine in a wind farm according to claim 6, characterized in that, The specific content in Step 4 is as follows: Step 4-1: Feature fusion processing. Perform feature fusion on the temporal features and spatial features of the wind speed of the wind farm units. The fusion process is as follows: Z = H + g θ * G X Step 4-2: Prediction output of the wind speed of each unit in the wind farm. Based on the fusion features and the output of the multi-layer fully connected layer, the prediction result is calculated. The calculation process is as follows: Y = W y2 (W y1 Z + b y1 ) + b y2 where W y1 , b y1 are the weights and biases of the first fully connected layer, W y2 , b y2 are the weights and biases of the second fully connected layer, and Y is the predicted result output.

Images