Wind speed prediction method and system based on hybrid convolutional network and parallel prediction model

Abstract

The invention discloses a wind speed prediction method and system based on a hybrid convolutional network and a parallel prediction model, and relates to the technical field of weather forecast, and the method comprises the steps: firstly carrying out the cluster screening through a clustering method, and then selecting a certain turbine as a target data set for prediction. During prediction, a GCN model is used to carry out spatial feature learning on historical wind speed data, then a TCN model is used to carry out time feature learning on the historical wind speed data, and then time sequence format data constructed by fusing spatial and temporal features are input into a Transform-LSTM parallel structure prediction model for prediction. According to the wind speed prediction method and system based on the hybrid convolutional network and the parallel prediction model, which adopt the structure and the steps, the prediction result is highly consistent with the observation value in numerical value, and relatively high accuracy is also shown in the aspect of trend prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and particularly to a wind speed prediction method and system based on a hybrid convolutional network and a parallel prediction model. Background Art

[0002] Wind speed is one of the important parameters in weather forecasting and also an important influencing factor for wind power generation. Through accurate wind speed prediction, the wind power resources in the future for a period of time can be predicted in advance, which has wide applications in aspects such as agricultural production management, disaster warning, aviation and navigation, and the performance and safety of structures. It is also of great significance for optimizing the operation and dispatch of wind turbines, reducing the backup and balancing costs of the power grid system, and increasing the grid-connected scale of wind power generation.

[0003] Existing wind speed prediction methods include physical methods, statistical methods, machine learning, deep learning, etc. These single models have many problems. Physical methods are based on the simulation of atmospheric physical processes, require high computing resources, and have limited prediction effects on short time scales; statistical methods mainly model based on the linear relationship of historical data and are difficult to handle the non-linear characteristics and abnormal changes in wind speed data; machine learning methods have limitations in dealing with the non-stationarity and multi-scale characteristics of wind speed data and cannot well take into account the complexity of time and space dimensions; deep learning methods are strongly data-dependent, the model performance depends on data quality and quantity, and they have high computational complexity and long training time. Summary of the Invention

[0004] The purpose of the present invention is to provide a wind speed prediction method and system based on a hybrid convolutional network and a parallel prediction model, and to improve the accuracy and efficiency of powder high-speed prediction by optimizing the model and calculation method.

[0005] To achieve the above purpose, the present invention provides a wind speed prediction method and system based on a hybrid convolutional network and a parallel prediction model, and the steps are as follows:

[0006] S1. Multi-dimensional data preprocessing: cleaning, filling, and normalizing the meteorological data of the wind farm and the operation data of the turbines;

[0007] S2. Clustering of wind farm turbine units: using the DBSCAN algorithm to divide the turbine units into clusters with similar operating characteristics;

[0008] S3. Spatiotemporal feature extraction: extracting spatial features through the graph convolutional neural network GCN and combining with the temporal convolutional network TCN to extract temporal features;

[0009] S4. Construction of parallel prediction model: fusing the output results of the Transformer model and the LSTM model to generate the final prediction result.

[0010] Preferably, S1 specifically includes:

[0011] S11. Collect the meteorological data and turbine data every 10 minutes through the monitoring system of the wind farm;

[0012] S12. Clean the collected meteorological data and turbine operation data, including improving data accuracy by setting thresholds and data verification rules, and filling missing data by interpolation method;

[0013] S13. Use the normalization method to eliminate the differences between different data dimensions and dimensions. The normalization formula is as follows:

[0014]

[0015] where X i , respectively represent the sequence elements before and after normalization; μ represents the mean of the sequence data, and σ represents the standard deviation of the sequence data.

[0016] Preferably, S2 specifically includes:

[0017] S21. Let the projection data points x i of the data corresponding to each turbine in the wind farm form a data set D = {x1,..., x n}, determine the neighborhood radius ε and the minimum number of points MinPts of the data set D, and calculate the neighborhood of each turbine data point: N ε (x i ) = {x j ∈ D | d(x i , x j ≤ ε)}, which represents the set of data points in the data set whose distance d from x i is less than or equal to ε. The distance is calculated using the cosine distance, and the calculation formula is:

[0018]

[0019] S22. Based on ε and the minimum number of points MinPts, divide the core points, boundary points and noise points, including:

[0020] For any x i ,

[0021] If |N ε (x i )| ≥ MinPts, then mark this point as a core point;

[0022] If |N ε (x i )| < MinPts and x i ∈ N ε (xj ) Then mark this point as a boundary point;

[0023] Otherwise, mark it as a noise point;

[0024] S23. Generate turbine clusters by expanding from core points, including:

[0025] Traverse all unvisited core points and create a new cluster C;

[0026] Add the current core point to cluster C and mark it as visited;

[0027] For the neighborhood N ε (x i ) of the current core point, for each point x j , if x j is not visited, mark it as visited. If x j is a core point and has not been assigned to any cluster, recursively visit its neighborhood;

[0028] Add all boundary points in the neighborhood to cluster C;

[0029] Repeat the steps of S23 until all points are processed.

[0030] Preferably, S3 specifically includes:

[0031] S31. The steps to construct a graph convolutional neural network GCN model are as follows:

[0032] (1) Construct a graph structure:

[0033] Map each turbine in the wind farm to a graph node, and define the edges between nodes based on geographical location, power transmission relationship, and correlation;

[0034] Construct an adjacency matrix A = {a ij}, n×n where n represents the number of nodes, and a ij represents the weight of the edge between node i and node j. When node i and node j are not connected, a ij = 0;

[0035] Calculate the degree matrix D. The degree matrix is a diagonal matrix, and the diagonal element is the degree of node i, representing the sum of the weights of all edges connected to node i;

[0036] (2) Construct a GCN network structure: The input layer of the GCN network is used to receive preprocessed data. The GCN network has multiple stacked graph convolutional layers, and the output of each graph convolutional layer is used as the input of the next layer for convolutional operations in multiple spatial dimensions;

[0037] The number of nodes in the input layer of the GCN network is n×d, where d is the dimension of the input data;

[0038] The graph convolutional layer of the GCN network is used to perform convolutional operations on graph data, which is expressed by the formula:

[0039]

[0040] where H l is the output feature matrix of the l-th layer, with dimension n×f l , and f l represents the feature dimension of the l-th layer;

[0041] is the adjacency matrix with self-connection added, and I is the identity matrix;

[0042] is 's degree matrix; W l is the weight matrix of the l-th layer, with dimension f l ×f l+1 ;

[0043] σ is the activation function, and the ReLU function is adopted:

[0044] ReLU(x) = max(0, x) (4)

[0045] S32. The steps to construct the temporal convolutional network TCN are as follows:

[0046] The convolutional kernel size of the convolutional layer of TCN is k, the number of input channels is c in , and the number of output channels is c out , and perform dilated convolutional operations on the output of the GCN network in the time dimension, which is expressed by the formula:

[0047] ]>

[0048] where y t represents the output of the convolutional layer at time step t, x t is the input or the output of the previous layer at time step t, d is the dilation rate, ω ij is the weight of the convolutional kernel, and b is the bias term;

[0049] After the convolutional operation, a residual connection is introduced to optimize the gradient propagation, which is expressed by the formula:

[0050] z = x + y (6)

[0051] The output z with the residual connection will be used as the input of the next layer after passing through the activation function σ for non-linear transformation.

[0052] Preferably, S4 specifically includes:

[0053] S41: Build the model of Transformer, use the output result of the Temporal Convolutional Network (TCN) as the input sequence, convert each element s in the input sequence into a low-dimensional vector representation, and embed the positional encoding (PE) to obtain the positional embedding vector s pos :

[0054] s pos = Embedding(s) + PE (7)

[0055] Calculate the attention weights Attention(Q, K, V) using the attention mechanism:

[0056]

[0057] where Q, K, and V represent query, key, and value respectively; <|

[0058] After concatenating the multi-head attention results, perform a linear transformation with the weight W O :

[0059] MultiHead(Q, K, V) = Concat(head1, …, head2)W O (9)

[0060] Use a feed-forward neural network to perform a non-linear transformation on the attention output:

[0061] FFN(x) = max(0, xW1 + b1)W2 + b2 (10)

[0062] x represents the input sequence at the current step and is also the output of the previous step,

[0063] Adopt layer normalization and residual connection LayerNorm(x + Sublayer(x)) to accelerate training;

[0064] When generating the output at the current position, perform a masking operation on the attention scores, calculate the attention using the encoder output, fuse the encoder information for the current decoding position, and perform a non-linear transformation through a feed-forward neural network;

[0065] Obtain the probability distribution of the elements based on the output of the decoder:

[0066] P(y|x) = softmax(Linear(DecoderOutput)) (11)

[0067] Select the maximum probability value as the prediction result;

[0068] S42: Build the model structure of the Long Short-Term Memory (LSTM) network

[0069] The number of neurons in the input layer of the LSTM network is equal to the dimension of the input features;

[0070] In the LSTM layer of the LSTM network,

[0071] The opening degree calculation formula of the forget gate is as follows:

[0072] f t = σ(W f ·[h t-1 , x t + b f ) (12)

[0073] Where x t is the input at the current time step, h t-1 is the hidden state at the previous time step, W f is the weight matrix of the forget gate, and b f is the bias of the forget gate;

[0074] The opening degree calculation formula of the input gate is as follows:

[0075] i t = σ(W i ·[h t-1 , x t + b i ) (13)

[0076] The calculation formula of the candidate cell state is as follows:

[0077]

[0078] Where i t is the output of the input gate at time step t, W i , W c are the weight matrices, and b i , b c are the biases;

[0079] The formula for updating the cell state is as follows:

[0080]

[0081] Where C t is the cell state at time step t, and ⊙ represents the Hadamard product;

[0082] The opening degree calculation formula of the output gate is as follows:

[0083] o t = σ(W o ·[h t-1 , x t + b o ) (16)

[0084] The calculation formula for the hidden state is as follows:

[0085] h t = o t ⊙ tanh(C t ) (17)

[0086] where W o is the weight matrix of the output gate, and b o is the bias;

[0087] S43: Fuse the outputs of the Transformer and LSTM, and output the prediction results of the GCN-TCN-Transformer-LSTM network structure through a linear layer.

[0088] A wind speed prediction system based on a hybrid convolutional network and a parallel prediction model includes:

[0089] A data acquisition module, including an internal data acquisition module for recording the operating status of the wind turbines and an external data acquisition module for recording meteorological information;

[0090] An information transmission module for uploading the recorded data to the central computing platform, pushing the program operation results to the interaction interface, and receiving interaction instructions;

[0091] An interaction module for displaying the prediction results, viewing the real-time operating status and meteorological information, and inputting instructions

[0092] A central computing platform for quickly processing the data uploaded by the acquisition module through a pre-written application program;

[0093] An application program stored on the central computing platform, developed based on the constructed GCN-TCN-Transformer-LSTM network structure, including a clustering module, a feature extraction module, and a parallel prediction module. The clustering module is used for clustering processing of the electric field turbine units, the feature extraction module is used for spatio-temporal feature extraction processing of the turbine unit clusters obtained by clustering according to groups, and the parallel prediction module is used for parallel prediction of the Transformer and LSTM according to the results of the spatio-temporal feature extraction.

[0094] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0095] First, multiple wind turbines are usually erected in a wind farm. Since the environments and meteorological conditions of adjacent wind turbines are highly similar, there is also a strong correlation in the wind speed changes between them. By using a spatio-temporal graph convolutional neural network, the time and space correlations across the wind farm can be effectively extracted, and this correlation can be used to improve the accuracy of wind speed prediction.

[0096] Secondly, a parallel structure combining Transformer and LSTM is adopted, which effectively improves the model's processing ability for complex data and prediction efficiency, meeting the real-time processing requirements of large-scale data. This method has obvious advantages in improving prediction accuracy. The system built based on this structure is easy to be connected to other functional systems for collaboration. For example, it can be connected to the wind farm operation control system to provide a basis for wind farm operation scheduling, helping to optimize the power scheduling of the wind farm, reduce energy waste, and has important practical application value, especially in the fields of smart grid and renewable energy with broad prospects.

[0097] The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0099] Figure 1 It is a schematic flowchart of the wind speed prediction method based on the hybrid convolutional network and the parallel prediction model according to the embodiment of the present invention;

[0100] Figure 2 It is a schematic architecture diagram of the wind speed prediction system based on the hybrid convolutional network and the parallel prediction model according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0101] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0102] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0103] Embodiment

[0104] As Figure 1 shown, the wind speed prediction method and system based on the hybrid convolutional network and the parallel prediction model are as follows:

[0105] S1. Multi-dimensional data preprocessing: Clean, fill, and normalize the meteorological data and turbine operation data of the wind farm; the data includes wind speed, wind direction, ambient temperature, nacelle temperature, nacelle direction, blade pitch angle, reactive power output, and active power output.

[0106] Specifically include:

[0107] S11. Through the monitoring system of the wind farm, collect the meteorological data and turbine data every 10 minutes;

[0108] S12. Clean the collected meteorological data and turbine operation data, including improving data accuracy by setting thresholds and data verification rules, identifying and removing outliers caused by equipment failures or transmission problems, and filling missing data by interpolation to avoid the impact of missing data on the prediction model;

[0109] S13. Use the normalization method to eliminate the differences between different data dimensions and measurement units. The normalization formula is as follows:

[0110]

[0111] Among them, X i , respectively represent the sequence elements before and after normalization; μ represents the mean of the sequence data, and σ represents the standard deviation of the sequence data.

[0112] S2. Clustering of wind farm turbine units: Use the DBSCAN algorithm to divide the turbine units into sub-turbine unit clusters with similar operating characteristics and strong regularity. Specifically include:

[0113] S21. Let the projection data points x of the corresponding data of each turbine in the wind farm in the high-dimensional space i constitute the data set D = {x1,..., x n}, determine the neighborhood radius ε and the minimum number of points MinPts of the data set D, and calculate the neighborhood of each turbine data point: N ε (x i ) = {x j ∈D|d(x i , x j ≤ε)}, which represents the set of data points in the data set whose distance d from x i is less than or equal to ε. Among them, the distance uses the cosine distance, and the calculation formula is:

[0114]

[0115] S22. Based on ε and the minimum number of points MinPts, divide the core points, boundary points, and noise points, including:

[0116] For any xi ,

[0117] If |N ε (x i )| ≥ MinPts, then mark this point as a core point;

[0118] If |N ε (x i )| < MinPts and x i ∈ N ε (x j ), then mark this point as a border point;

[0119] Otherwise, mark it as a noise point;

[0120] S23. Generate turbine unit clusters by expanding through core points, including:

[0121] Traverse all unvisited core points and create a new cluster C;

[0122] Add the current core point to cluster C and mark it as visited;

[0123] For each point x ε in the neighborhood N i (x j ), if x j is not visited, then mark it as visited. If x j is a core point and has not been assigned to any cluster, then recursively visit its neighborhood;

[0124] Add all border points in the neighborhood to cluster C;

[0125] Repeat the steps of S23 until all points are processed. Through clustering, multiple turbine unit clusters with similar characteristics in the wind farm are obtained. Then, model training and prediction are carried out separately according to the obtained clusters.

[0126] The GCN-TCN-Transformer-LSTM constructed in the present invention is the main prediction model, where GCN-TCN is the feature extraction model and Transformer-LSTM is the parallel prediction model.

[0127] S3. Spatiotemporal feature extraction: Extract spatial features through the graph convolutional neural network GCN and combine with the temporal convolutional network TCN to extract temporal features; specifically including:

[0128] S31. The steps to construct the graph convolutional neural network GCN model are as follows:

[0129] (1) Construct a graph structure:

[0130] Map each turbine in the wind farm into a graph node and define the edges between nodes based on their geographical location, power transmission relationship and correlation;

[0131] Construct the adjacency matrix A={a ij} n×n , where n represents the number of nodes, a ij Represents the weight of the edge between node i and node j. When node i and node j are not connected, a ij =0;

[0132] Calculate the degree matrix D. The degree matrix is a diagonal matrix with diagonal elements is the degree of node i, which represents the sum of the weights of all edges connected to node i;

[0133] (2) Constructing the GCN network structure: The input layer of the GCN network is used to receive preprocessed data. In the GCN network, multiple graph convolution layers can be stacked based on the actual data to learn more complex graph structures and features. The output of each graph convolution layer serves as the input of the next layer, performing convolution operations on multiple spatial dimensions.

[0134] The number of nodes in the input layer of the GCN network is n×d, where d is the dimension of the input data;

[0135] The graph convolution layer of the GCN network is used to perform convolution operations on graph data, which can be expressed as follows:

[0136]

[0137] Among them, H l is the output feature matrix of the lth layer, with dimension n×f l , f l Represents the feature dimension of the lth layer;

[0138] is the adjacency matrix with self-connection added, and I is the identity matrix;

[0139] yes The degree matrix of l is the weight matrix of layer l, with dimension f l ×f l+1 ;

[0140] σ is the activation function, using the ReLU function:

[0141] ReLU(x)=max(0,x) (4)

[0142] S32. The steps to construct the temporal convolutional network TCN are as follows:

[0143] The convolution kernel size of TCN is k, and the number of input channels is c.in The number of output channels is c out Perform dilated convolution operation on the output of the GCN network in the time dimension, which is expressed by the formula:

[0144]

[0145] where y t represents the output of the convolutional layer at time step t, and x t is the output of the input or the previous layer at time step t, d is the dilation rate, ω ij is the weight of the convolutional kernel, and b is the bias term;

[0146] Introduce residual connection after the convolution operation to optimize gradient propagation, which is expressed by the formula:

[0147] z = x + y (6)

[0148] The output with the residual connection z will be used as the input of the next layer after passing through the activation function σ for non-linear transformation. According to the specific task, choose whether to use a pooling layer to downsample the time series to reduce the data dimension and extract more important features.

[0149] S4. Construction of the parallel prediction model: Fuse the output results of the Transformer model and the LSTM model to generate the final prediction result. Specifically, it includes:

[0150] S41: Construct the Transformer model, use the output result of the time convolutional network TCN as the input sequence, convert each element s in the input sequence into a low-dimensional vector representation, and embed the position encoding PE to obtain the position embedding vector s pos :

[0151] s pos = Embedding(s) + PE (7)

[0152] Calculate the attention weight Attention(Q, K, V) using the attention mechanism:

[0153]

[0154] where Q, K, and V represent the query, key, and value respectively;

[0155] After concatenating the multi-head attention results, perform a linear transformation with the weight W O :

[0156] MultiHead(Q, K, V) = Concat(head1,..., head2)W O (9)

[0157] Use a feedforward neural network to perform a nonlinear transformation on the attention output:

[0158] FFN(x)=max(0,xW1+b1)W2+b2 (10)

[0159] x represents the input sequence of the current step and is also the output of the previous step.

[0160] To prevent the gradient from disappearing or exploding, layer normalization and residual connection LayerNorm(x+Sublayer(x)) are used to accelerate training;

[0161] The attention score is masked when generating the output of the current position to prevent the model from leaking information about future sequences. The attention is calculated using the encoder output, and the encoder information is fused for the current decoding position, and a nonlinear transformation is performed through a feedforward neural network;

[0162] The probability distribution of the elements is obtained according to the output of the decoder:

[0163] P(y|x)=softmax(Linear(DecoderOutput)) (11)

[0164] Select the maximum probability value as the prediction result;

[0165] S42: Construct the model structure of the long short-term memory network LSTM,

[0166] The number of neurons in the input layer of the LSTM network is equal to the dimension of the input features;

[0167] In the LSTM layer of the LSTM network,

[0168] The forget gate determines which information is discarded from the cell state. The formula for calculating the opening of the forget gate is as follows:

[0169] f t =σ(W f ·[h t-1 ,x t ]+b f ) (12)

[0170] where x t is the input at the current moment, h t-1 is the hidden state of the previous moment, W f is the weight matrix of the forget gate, b f is the bias of the forget gate;

[0171] The input gate determines what new information is stored in the cell state. The opening formula of the input gate is as follows:

[0172] i t =σ(Wi · [h t-1 , x t + b i ) (13)

[0173] The calculation formula for the candidate cell state is as follows:

[0174]

[0175] where i t is the output of the input gate at time t, W i , W c is the weight matrix, b i , b c is the bias;

[0176] The formula for updating the cell state is as follows:

[0177]

[0178] where C t is the cell state at time t, and ⊙ represents the Hadamard product;

[0179] The calculation formula for the opening degree of the output gate is as follows:

[0180] o t = σ(W o · [h t-1 , x t + b o ) (16)

[0181] The calculation formula for the hidden state is as follows:

[0182] h t = o t ⊙ tanh(C t ) (17)

[0183] where W o is the weight matrix of the output gate, b o is the bias;

[0184] S43: Fuse the outputs of the Transformer and LSTM, and output the prediction result through a linear layer.

[0185] The above model first inputs the historical monitoring data into the constructed network architecture, and trains and updates the network parameters through the backpropagation algorithm. After the training is completed, the sequence to be predicted is input into the trained GCN-TCN-Transformer-LSTM network for prediction.

[0186] To execute the above prediction method, the present invention proposes a wind speed prediction system based on a hybrid convolutional network and a parallel prediction model, asFigure 2 As shown in the figure, it includes:

[0187] A data acquisition module, including an internal data acquisition module for recording the operating status of the fan and an external data acquisition module for recording meteorological information;

[0188] An information transmission module, used to upload the recorded data to the central computing platform, push the program operation results to the interaction interface, and receive interaction instructions;

[0189] An interaction module, used to display the prediction results, view the real-time operating status and meteorological information, and input instructions

[0190] A central computing platform, which quickly processes the data uploaded by the acquisition module through a pre-written application program;

[0191] The application program stored on the central computing platform is developed based on the constructed GCN-TCN-Transformer-LSTM network structure, including a clustering module, a feature extraction module, and a parallel prediction module. The clustering module is used for clustering processing of the electric field turbine units, the feature extraction module is used for spatio-temporal feature extraction processing of the turbine unit clusters obtained by clustering according to groups, and the parallel prediction module is used for parallel prediction of Transformer and LSTM according to the results of spatio-temporal feature extraction.

[0192] In order to verify the model performance of the technical solution of this application, an index evaluation based on error analysis is carried out on it.

[0193] The evaluation indexes of the model prediction performance usually adopt the Mean Absolute Error (MAE), the Mean-Square Error (MSE), and the Root Mean Square Error (RMSE). The calculation formulas are as follows:

[0194]

[0195] Taking the wind farm data in a certain place in Inner Mongolia as an example for analysis. It can be seen from the comparison of the following table results that the prediction result of the present invention is the best.

[0196]

[0197] In summary, the present invention proposes a wind speed prediction method and system based on Graph Convolutional Network (GCN) - Temporal Convolutional Network (TCN) and Transformer - Long Short - Term Memory (LSTM). The graph convolutional network can capture the spatial relationships between geographical units, make full use of the surrounding meteorological information, and solve the problem of spatial heterogeneity; TCN can extract local features of time series and expand the receptive field to handle long - term dependence relationships; the self - attention mechanism of Transformer is good at capturing long - range dependencies and learning the long - term patterns of wind speed data; LSTM can effectively handle the long - term dependence and gradient disappearance problems of time series, and combined with the Transformer model can further improve the ability to capture long - term dependencies of wind speed data. These improvements help to improve the accuracy and efficiency of wind speed prediction and overcome the deficiencies in the prior art.

[0198] For the remaining technical features in the above - mentioned embodiments, those skilled in the art can flexibly select them according to the actual situation to meet different specific actual needs. However, it is obvious to those of ordinary skill in the art that these specific details do not have to be adopted to implement the present invention. In other instances, in order to avoid confusing the present invention, well - known components, structures, or parts are not specifically described, and all are within the scope of the technical solutions claimed in the claims of the present invention.

[0199] Any modifications and changes made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention. In the above description, in order to provide a thorough understanding of the present invention, a large number of specific details are elaborated. However, it is obvious to those of ordinary skill in the art that these specific details do not have to be adopted to implement the present invention. In other instances, in order to avoid confusing the present invention, well - known technologies, such as specific construction details, operating conditions, and other technical conditions, are not specifically described.

[0200] In this article, specific examples are used to elaborate on the principle and implementation mode of the present invention. The description of the above - mentioned embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation mode and application scope. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A wind speed prediction method and system based on a hybrid convolutional network and a parallel prediction model, characterized in that: The steps are as follows: S1. Multi-dimensional data preprocessing: cleaning, filling, and normalizing the meteorological data and turbine operation data of the wind farm; S2. Clustering of wind farm turbine units: using the DBSCAN algorithm to divide the turbine units into clusters with similar operation characteristics; S3. Spatiotemporal feature extraction: extracting spatial features through the graph convolutional neural network GCN and combining with the temporal convolutional network TCN to extract temporal features; S4. Construction of a parallel prediction model: fusing the output results of the Transformer model and the LSTM model to generate the final prediction result.

2. The wind speed prediction method based on a hybrid convolutional network and a parallel prediction model according to claim 1, characterized in that: S1 specifically includes: S11. Through the monitoring system of the wind farm, collect meteorological data and turbine data every 10 minutes; S12. Clean the collected meteorological data and turbine operation data, including improving data accuracy by setting thresholds and data verification rules, and filling missing data by interpolation; S13. Use the normalization method to eliminate the differences between different data dimensions and measurement units. The normalization formula is as follows: Among them, respectively represent the sequence elements before and after normalization; μ represents the mean of the sequence data, and σ represents the standard deviation of the sequence data.

3. The wind speed prediction method based on a hybrid convolutional network and a parallel prediction model according to claim 2, characterized in that: S2 specifically includes: S21. Let the projection data points \(x\) of the corresponding data of each turbine in the wind farm in the high-dimensional space i constitute the data set \(D = \{x_1, \ldots, x\) n \}. Determine the neighborhood radius \(\varepsilon\) and the minimum number of points MinPts of the data set \(D\), and calculate the neighborhood of each turbine data point: \(N\) ε (x i ) = \{x j \in D|d(x i , x j \leq\varepsilon\}\), which represents the set of data points in the data set whose distance \(d\) from \(x i is less than or equal to \(v\). The distance is calculated using the cosine distance, and the calculation formula is: S22. Divide core points, border points, and noise points based on ε and the minimum number of points MinPts, including: For any x i , If |N ε (x i )| ≥ MinPts, then mark this point as a core point; If |N ε (x i )| < MinPts and x i ∈ N ε (x j ), then mark this point as a boundary point; Otherwise, mark it as a noise point; S23. Generate turbine unit clusters by expanding core points, including: Traverse all unvisited core points and create new clusters; Add the current core point to cluster C and mark it as visited; For the neighborhood N of the current core point ε (x i ) for each point x j in it, if x j has not been visited, mark it as visited. If x j is a core point and has not been assigned to any cluster, recursively visit its neighborhood; Add all border points in the neighborhood to cluster C; Repeat the steps of S23 until all points are processed.

4. The wind speed prediction method based on a hybrid convolutional network and a parallel prediction model according to claim 3, characterized in that: S3 specifically includes: The steps to construct the graph convolutional neural network GCN model are as follows: (1) Construct the graph structure: Map each turbine in the wind farm to a graph node, and define the edges between nodes based on geographical location, power transmission relationship, and correlation; Construct the adjacency matrix A = {a ij} n×n , where n represents the number of nodes, and a ij represents the weight of the edge between node i and node j. When node i and node j are not connected, a ij = 0; Calculate the degree matrix D, which is a diagonal matrix, and the diagonal elements are the degrees of node i, representing the sum of the weights of all edges connected to node i; (2) Construct the GCN network structure: The input layer of the GCN network is used to receive the preprocessed data. The GCN network has multiple stacked graph convolutional layers, and the output of each graph convolutional layer is used as the input of the next layer for convolutional operations in multiple spatial dimensions; The number of nodes in the input layer of the GCN network is n×d, where d is the dimension of the input data; The graph convolutional layer of the GCN network is used to perform convolutional operations on graph data, expressed by the formula: Among them, H l is the output feature matrix of the l-th layer, with a dimension of n×f l , f l represents the feature dimension of the l-th layer; is the adjacency matrix with self-connections added, and I is the identity matrix; is the degree matrix; W l is the weight matrix of the l-th layer, with dimension f l ×f l+1 ; σ is the activation function, using the ReLU function: ReLU(x) = max(0, x) (4) The steps to construct the temporal convolutional network TCN are as follows: The convolution kernel size k of the convolutional layer of TCN, and the number of input channels is c in , and the number of output channels is c out , perform dilated convolution operation on the output of the GCN network in the time dimension, and the formula is expressed as: Among them, y t represents the output of the convolutional layer at time step t, and x t is the input or the output of the previous layer at time step t, d is the dilation rate, and ω ij is the weight of the convolutional kernel, and b is the bias term; Introduce a residual connection after the convolutional operation to optimize gradient propagation, expressed by the formula: z = x + y (6) The output z with the residual connection will be used as the input of the next layer after passing through the activation function σ for nonlinear transformation.

5. The wind speed prediction method based on a hybrid convolutional network and a parallel prediction model according to claim 4, characterized in that: S4 specifically includes: S41: Build a Transformer model, use the output result of the temporal convolutional network TCN as the input sequence, convert each element s in the input sequence into a low-dimensional vector representation, and embed the positional encoding PE to obtain the positional embedding vector s pos : s pos = Embedding(s) + PE(7) Calculate the attention weight Attention(Q, K, V) using the attention mechanism: Among them, Q, K, and V represent the query, key, and value respectively; After concatenating the multi-head attention results, perform a linear transformation with weight W O : MultiHead(Q,K,V) = Concat(head1,…,head2)W O (9) Use a feed-forward neural network to perform nonlinear transformation on the attention output: FFN(x) = max(0, xW1 + b1)W2 + b2 (10) x represents the input sequence of the current step, which is also the output of the previous step, Adopt layer normalization and residual connection LayerNorm(x + Sublayer(x)) to accelerate training; Mask the attention scores when generating the current position output, calculate the attention using the encoder output, fuse the encoder information for the current decoding position, and perform a non-linear transformation through a feed-forward neural network; Obtain the probability distribution of elements according to the output of the decoder: P(y|x) = softmax(Linear(DecoderOutput)) (11) Select the maximum probability value as the prediction result; S42: Construct the model structure of the long short-term memory network LSTM, The number of neurons in the input layer of the LSTM network is equal to the dimension of the input features; In the LSTM layer of the LSTM network, The opening formula of the forget gate is as follows: f t = σ(W f · [h t-1 , x t + b f ) (12) where x t is the input at the current time, h t-1 is the hidden state at the previous time, W f is the weight matrix of the forget gate, b f is the bias of the forget gate; The opening formula of the input gate is as follows: i t = σ(W i · [h t-1 , x t + b i ) (13) The calculation formula of the candidate cell state is as follows: where i t is the output of the input gate at time t, W i , W c is the weight matrix, b i , b c is the bias; The formula for updating the cell state is as follows: Among which C t is the cell state at time t, and ⊙ represents the Hadamard product; The opening formula of the output gate is as follows: o t = σ(W o · [h t-1 , x t + b o ) (16) The calculation formula of the hidden state is as follows: h t = o t ⊙tanh(C t ) (17) where W o is the weight matrix of the output gate, and b o is the bias; S43: Fuse the outputs of the Transformer and LSTM, and output the prediction result of the GCN-TCN-Transformer-LSTM network structure through a linear layer.

6. A prediction system applying the wind speed prediction method based on a hybrid convolutional network and a parallel prediction model according to any one of claims 1-5, characterized in that, Including: The data acquisition module, including an internal data acquisition module for recording the operating state of the fan and an external data acquisition module for recording meteorological information; The information transmission module is used to upload the recorded data to the central computing platform, push the program operation result to the interaction interface, and receive interaction instructions; The interaction module is used to display the prediction result, view the real-time operating state and meteorological information, input instructions to the central computing platform, and quickly process the data uploaded by the acquisition module through a pre-written application program; The application program stored on the central computing platform is developed based on the constructed GCN-TCN-Transformer-LSTM network structure, including a clustering module, a feature extraction module, and a parallel prediction module. The clustering module is used to perform clustering processing on the electric field turbine units, the feature extraction module is used to perform spatio-temporal feature extraction processing on the turbine unit clusters obtained by clustering according to groups, and the parallel prediction module is used to perform parallel prediction of the Transformer and LSTM according to the results of spatio-temporal feature extraction.

Images