A Method and System for Predicting Offshore Wind Power with Multi-Level Spatiotemporal Awareness

Abstract

The present invention provides a method and system for predicting the power of offshore wind turbines with multi-level spatio-temporal perception, which relates to the technical field of predicting the power of offshore wind turbines. First, the spatial relationship of the wind turbines is determined, a two-dimensional static plane feature is constructed, multi-dimensional time-series wind power meteorological features at multiple altitude levels are obtained, the multi-dimensional time-series wind power meteorological features at each altitude level are dimensionally reduced, weights are assigned to the dimensionally reduced multi-dimensional time-series wind power meteorological features, a feature matrix after dimensional reduction at each altitude level is obtained, local graph features are constructed based on the feature matrix, and the local graph features and the two-dimensional static plane features are combined into global spatial graph features. Finally, a linear regression prediction model is used to predict the wind power. The present invention makes full use of the geographical location information and meteorological information of the wind turbines in the wind farm, and at the same time considers various meteorological factors at different altitude levels, effectively reducing the prediction error caused by the uncertainty of wind power generation and improving the accuracy of predicting the power of offshore wind turbines.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore wind power prediction, and more specifically, to an offshore wind power prediction method and system with multi-level spatio-temporal perception. Background Art

[0002] As an important direction for the development of clean energy, offshore wind power is developing rapidly and becoming an important part of the power system due to its advantages of rich and stable wind resources, high wind speed, and little impact on terrain, as well as the clean, rich, and environmentally friendly characteristics of wind energy itself. Compared with onshore wind farms, offshore wind farms have higher wind energy utilization efficiency, with an annual average utilization hours about 500 hours more than onshore wind power. However, the output of offshore wind power is significantly affected by natural conditions, especially the changes in wind speed, wind direction, and weather conditions. This uncertainty makes offshore wind power prediction challenging and also poses challenges to the stable operation of the power grid and power dispatching, directly affecting the economic benefits and operation efficiency of wind farms. Therefore, accurate offshore wind power prediction is of great significance for the safe and stable operation of the power system.

[0003] In the prior art, an offshore wind power prediction method based on RCNN and meteorological time series features is proposed. First, historical numerical weather forecasts and historical output power data of the wind farm are extracted, and the historical data is evenly divided into a training set and a validation set. An RCNN is constructed as a preliminary prediction model, and the RCNN is trained using the training set to optimize the weights. Secondly, the trained RCNN is used to predict the wind power in the validation set, and the difference between the RCNN prediction result and the historical wind power in the validation set is calculated to obtain the historical prediction power error. Then, the Xgboost algorithm is used to calculate the correlation between the numerical weather forecast in the validation set and the historical prediction power error, and the numerical weather forecasts with strong correlation are selected as meteorological time series features. Then, a bidirectional recurrent neural network is constructed, and the network is trained using the meteorological time series features and the historical prediction power error. Finally, the prediction results and errors of the RCNN and the bidirectional recurrent neural network are superimposed to obtain the offshore wind power prediction, which improves the overall prediction performance of the model while ensuring timeliness. However, this method only relies on the historical power sequence of offshore wind power for prediction, without considering the influence of the geographical location information of the wind turbines and the complex and changeable meteorological factors at sea on offshore wind power generation, resulting in poor generalization of the finally trained wind power prediction model and low prediction accuracy for offshore wind power under complex conditions.

[0004] In practical applications, the installation positions of wind turbines in a wind farm in a certain area will have an impact on the power generation of surrounding wind turbines. Therefore, the geographical location information of the wind turbines themselves is also an important feature that cannot be ignored. The meteorological factors at sea are complex and changeable, especially at high altitudes, where there are more significant changes. Considering various meteorological factors at different altitudes helps to improve the prediction accuracy. Summary of the Invention

[0005] To solve the problem of low prediction accuracy in the current method for predicting the power of offshore wind farms, the present invention proposes a method and system for predicting the power of offshore wind farms with multi-level spatio-temporal perception, reducing the prediction error caused by the uncertainty of wind power generation and improving the prediction accuracy of the wind power of offshore wind farms.

[0006] To achieve the above technical effects, the technical solution of the present invention is as follows:

[0007] In the first aspect, the present application proposes a method for predicting the power of offshore wind farms with multi-level spatio-temporal perception, including the following steps:

[0008] S1: Regarding each wind turbine in the offshore wind farm within the region as a node, determining the spatial relationship of the wind turbines according to the geographical location information of each wind turbine and the distance between the wind turbines, and constructing two-dimensional static plane features;

[0009] S2: Obtaining multi-dimensional time-series wind power meteorological features at multiple altitude levels, and performing dimensionality reduction on the multi-dimensional time-series wind power meteorological features at each altitude level;

[0010] S3: Assigning weights to the dimensionality-reduced multi-dimensional time-series wind power meteorological features, obtaining a feature matrix after dimensionality reduction at each altitude level, constructing local graph features based on the feature matrix after dimensionality reduction, and combining the local graph features and the two-dimensional static plane features into a global spatial graph feature;

[0011] S4: Inputting the global spatial graph feature into a graph convolutional neural network model to generate a final latent feature vector;

[0012] S5: Using the final latent feature vector to train the constructed linear regression prediction model, and using the trained linear regression prediction model for predicting the wind power of the offshore wind farm.

[0013] Preferably, the geographical location information of each wind turbine includes: the longitude, latitude and altitude of each wind turbine; representing the geographical information of the wind turbines in the offshore wind farm within the region as , where P represents the number of wind turbines, M represents the longitude, latitude and altitude of the wind turbine; represents the set of geographical information of the wind turbines;

[0014] The process of constructing two-dimensional static plane features is as follows:

[0015] S11: Calculating the normalized distance ND between the wind turbine i and the wind turbine j in the offshore wind farm within the region, and the calculation process is as follows:

[0016]

[0017]

[0018]

[0019]

[0020]

[0021] Among them, represents the longitude of the wind turbine i and represents the latitude of the wind turbine i and represents the altitude of the wind turbine i ; () is an intermediate operation function. For any hav , x , where R represents the radius of the earth, represents the distance between the wind turbine i and the wind turbine j ; ;

[0022] S12: Express the wind power i of the wind turbine as I , and express the wind power j of the wind turbine as J . Calculate the normalized mutual information NMI between the wind turbine i and the wind turbine j . The expression is:

[0023]

[0024]

[0025]

[0026]

[0027] Among them, is the joint distribution of the wind turbine i and the wind turbine j , is the marginal distribution of the wind turbine i , represents the information entropy of the wind turbine i , represents the information entropy of the wind turbine j , represents the information entropy of the wind turbine i and the wind turbine jThe joint information entropy, represents the fan i and the fan j mutual information;

[0028] S13: Construct a two-dimensional static plane feature based on the normalized distance and normalized mutual information between fan pairs .

[0029] Preferably, collect meteorological data in the acquisition area, and the meteorological data includes: wind speed V , wind direction D , temperature T , humidity H and air pressure P a , and the meteorological data changes with time, and preprocess the meteorological data;

[0030] Divide the altitude of the offshore wind farm in the area L into n levels, expressed as: , the i th level L i The altitude range of the altitude z is fixed and satisfies: , where represents the lower limit of the altitude range of the i th level L i , represents the upper limit of the altitude range of the i th level L i ;

[0031] At each altitude z, the meteorological data feature M ( z ) is expressed as a multi-dimensional vector, and the expression is:

[0032] ;

[0033] The i th level L i The meteorological data feature M ( L i ) is expressed as:

[0034] .

[0035] Preferably, the process of dimensionality reduction for the multi-dimensional time-series wind power meteorological features at each level of altitude is:

[0036] S21: Set a fixed sliding window Q, and select Q consecutive sampling points from the time series to form a subsequence , where t is the current time point, x represents any one of the meteorological variables, and the meteorological variables include wind speed V , wind direction D , temperature T , humidity H and air pressure P a ;

[0037] S22: On each subsequence q , construct a time series feature vector for each meteorological variable separately :

[0038]

[0039] Among them, each row represents the time series features of a single meteorological variable of a single node within each hierarchical altitude. represents the time series mean of the collection point q, represents the time series variance of the collection point q, represents the time autocorrelation coefficient of the collection point q, represents the spatial autocorrelation of the collection point q; all meteorological feature vectors of the node u within each hierarchical altitude are:

[0040]

[0041] Integrate all node features to obtain all node feature matrices f within each hierarchical altitude:

[0042]

[0043] S23: Perform principal component analysis on the feature matrix f to obtain a new feature matrix after dimensionality reduction . Calculate the covariance matrix C for and perform eigenvalue decomposition on it to obtain eigenvalues and corresponding eigenvectors and . Select the eigenvectors corresponding to the top k largest eigenvalues to form a projection matrix W, and obtain the expression of the feature matrix after dimensionality reduction within each hierarchical altitude as:

[0044]

[0045] Among them, has a size of n * k , 2 < k < mn is the number of meteorological variables, m is the total number of eigenvalues after decomposition.

[0046] Preferably, , , and The calculation expressions of are respectively:

[0047]

[0048]

[0049]

[0050]

[0051] Among them, is the spatial weight between the position of the fan i and the position of the fan j .

[0052] Preferably, the process of step S3 is as follows:

[0053] S31: Adopt the method of dynamic feature selection to set the weight allocation strategy, so that the weight of the feature matrix in the high-altitude layer is larger, and the sum of all weights is 1; the weight allocation strategy for a single-layer altitude layer is:

[0054]

[0055] Among them, z is the height of the L i th altitude layer, is an adjustable parameter used to control the sensitivity of the weight to different altitudes, and to achieve the control of the difference sensitivity between high altitude and low altitude, is the growth index used to amplify the weight of the high-altitude layer, is the smoothing control coefficient;

[0056] Through weighted summation, the feature matrices after dimensionality reduction of each altitude layer are merged into a local feature matrix :

[0057]

[0058] Among them, N is the number of single-layer altitude layers, is the weight of the th altitude layer;

[0059] S32: Use a multi-layer perceptron (MLP) to expand the dimension of the two-dimensional static plane features so that the dimension of the two-dimensional static plane features is the same as that of the local feature matrix After the dimension expansion, the two-dimensional static plane feature matrix and the local feature matrix are vertically concatenated to form the global spatial graph features F .

[0060] Preferably, the process of using the multi-layer perceptron (MLP) to expand the dimension of the two-dimensional static plane features is as follows:

[0061] Preprocess the two-dimensional static plane features ;

[0062] Input the preprocessed two-dimensional static plane features into the trained multi-layer perceptron (MLP);

[0063] Calculate the one-dimensional vector output by the output layer of the multi-layer perceptron through the forward propagation algorithm of the multi-layer perceptron (MLP);

[0064] Convert the one-dimensional vector output by the output layer into a matrix form with the same dimension as the local feature matrix .

[0065] Preferably, the process of generating the final latent feature vector described in step S4 is as follows:

[0066] S41: Construct the adjacency matrix A F according to the cosine similarity between the fan nodes in the global spatial graph features h , and the expression is:

[0067] ;

[0068] S42: Perform a forward convolution operation on the global spatial graph features F and the adjacency matrix A h in the graph convolutional neural network model with several layers; among them, the convolution operation in any layer of the graph convolutional neural network model satisfies:

[0069]

[0070] where E is the degree matrix, is the weight matrix of the first layer of the graph convolutional neural network model layer, is the feature matrix after the first layer of graph convolution operation, is the pThe weight matrix of the layer graph convolution; after performing the multi-layer graph convolution operation, extract the output features of the last layer of the graph convolutional neural network model, and the output features include the final potential feature vector of the wind turbines in the offshore wind farm.

[0071] Preferably, the process of step S5 is as follows:

[0072] Use the potential feature vector and the corresponding historical wind power data as training data, and train the linear regression prediction model through forward propagation;

[0073] Calculate the prediction error using the loss function, update the model parameter weights of the linear regression prediction model using the optimizer, and end the training after the loss function converges or reaches the predetermined number of training epochs to obtain the trained linear regression prediction model;

[0074] Execute steps S1 to S4 to obtain a new potential feature vector, input the new potential feature vector into the trained linear regression prediction model, and output the wind power of the offshore wind farm.

[0075] In a second aspect, the present application proposes an offshore wind power prediction system with multi-level spatio-temporal perception, characterized in that the system is used to implement the offshore wind power prediction method with multi-level spatio-temporal perception, including:

[0076] A two-dimensional static plane feature module, which regards each wind turbine in the offshore wind farm within the area as a node, determines the spatial relationship of the wind turbines according to the geographical location information of each wind turbine and the distance between the wind turbines, and constructs two-dimensional static plane features;

[0077] A dimensionality reduction module, which obtains multi-dimensional time-series wind power meteorological features at multiple altitude levels and reduces the dimensionality of the multi-dimensional time-series wind power meteorological features at each altitude level;

[0078] A feature integration and merging module, which is used to assign weights to the dimensionality-reduced multi-dimensional time-series wind power meteorological features, obtain the feature matrix after dimensionality reduction at each altitude level, construct local graph features based on the feature matrix after dimensionality reduction, and merge the local graph features and the two-dimensional static plane features into a global spatial graph feature;

[0079] An extraction module, which inputs the global spatial graph feature into the graph convolutional neural network model to generate the final potential feature vector;

[0080] A prediction module, which uses the final potential feature vector to train the pre-constructed linear regression prediction model, and uses the trained linear regression prediction model for the wind power prediction of the offshore wind farm.

[0081] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0082] The present invention provides a method and system for predicting the power of offshore wind turbines with multi-level spatio-temporal perception. First, based on the geographical location information of each wind turbine and the distance between wind turbines, the spatial relationship of the wind turbines is determined, and a two-dimensional static plane feature is constructed. Then, multi-dimensional time-series wind power meteorological features at multiple altitude levels are obtained to capture the meteorological features of different altitude layers changing over time. To reduce data redundancy and highlight key features, the present invention reduces the dimension of the multi-dimensional time-series wind power meteorological features at each altitude level, compresses the complexity of high-dimensional data, and retains the main changing trends. Weights are assigned to the multi-dimensional time-series wind power meteorological features after dimensionality reduction to obtain a feature matrix after dimensionality reduction at each altitude level. Based on the feature matrix after dimensionality reduction, local graph features are constructed, and the local graph features and the two-dimensional static plane features are combined into a global spatial graph feature. Finally, the global spatial graph feature is input into a graph convolutional neural network model to generate a final latent feature vector, and a linear regression prediction model is used to predict the wind power. The present invention makes full use of the geographical location information and meteorological information of the wind turbines in the wind farm, and at the same time considers various meteorological factors at different altitude levels, effectively reducing the prediction error caused by the uncertainty of wind power generation and improving the accuracy of offshore wind power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 FIG. shows a schematic flow chart of a method for predicting the power of offshore wind turbines with multi-level spatio-temporal perception proposed by an embodiment of the present invention;

[0084] Figure 2 FIG. shows an effect diagram of wind power prediction using the method for predicting the power of offshore wind turbines with multi-level spatio-temporal perception proposed by an embodiment of the present invention;

[0085] Figure 3 FIG. shows a structural diagram of a system for predicting the power of offshore wind turbines with multi-level spatio-temporal perception proposed by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0086] The drawings are only for illustrative purposes and should not be construed as a limitation to this application;

[0087] For better illustration of this embodiment, some parts of the drawings are omitted, enlarged or reduced, and do not represent the actual size;

[0088] For those skilled in the art, it is understandable that some well-known content descriptions in the drawings may be omitted.

[0089] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0090] The description of the positional relationship in the drawings is only for illustrative purposes and should not be construed as a limitation to this application;

[0091] Embodiment 1

[0092] This embodiment proposes a method for predicting the power of offshore wind farms with multi-level spatio-temporal perception. The flowchart of the implementation process of this method can be seen in Figure 1 , as Figure 1 shown, and it includes the following steps:

[0093] S1: Consider each wind turbine in the offshore wind farm within the region as a node, and determine the spatial relationship of the wind turbines according to the geographical location information of each wind turbine and the distance between the wind turbines, and construct two-dimensional static plane features;

[0094] S2: Obtain multi-dimensional time-series wind power meteorological features at multi-level altitudes, and reduce the dimension of the multi-dimensional time-series wind power meteorological features at each level of altitude;

[0095] S3: Assign weights to the dimension-reduced multi-dimensional time-series wind power meteorological features to obtain a feature matrix after dimension reduction at each level of altitude, construct local graph features based on the feature matrix after dimension reduction, and combine the local graph features and the two-dimensional static plane features into a global spatial graph feature;

[0096] S4: Input the global spatial graph feature into the graph convolutional neural network model to generate the final latent feature vector;

[0097] S5: Use the final latent feature vector to train the constructed linear regression prediction model, and use the trained linear regression prediction model for predicting the wind power of the offshore wind farm.

[0098] In this embodiment, first, according to the geographical location information of each wind turbine and the distance between the wind turbines, the spatial relationship of the wind turbines is determined, and two-dimensional static plane features are constructed. Then, multi-dimensional time-series wind power meteorological features at multi-level altitudes are obtained to capture the meteorological features of different altitude layers changing with time. To reduce data redundancy and highlight key features, the present invention reduces the dimension of the multi-dimensional time-series wind power meteorological features at each level of altitude, compresses the complexity of high-dimensional data, and retains the main changing trends. Weights are assigned to the dimension-reduced multi-dimensional time-series wind power meteorological features to obtain a feature matrix after dimension reduction at each level of altitude. Local graph features are constructed based on the feature matrix after dimension reduction, and the local graph features and the two-dimensional static plane features are combined into a global spatial graph feature. Finally, the global spatial graph feature is input into the graph convolutional neural network model to generate the final latent feature vector, and the linear regression prediction model is used for wind power prediction. The present invention makes full use of the geographical location information and meteorological information of the wind turbines in the wind farm, and at the same time considers various meteorological factors at different altitudes, effectively reducing the prediction error caused by the uncertainty of wind power generation and improving the accuracy of offshore wind power prediction.

[0099] Embodiment 2

[0100] In this embodiment, the geographical location information of each wind turbine includes the longitude, latitude, and altitude of each wind turbine; the geographical information of the wind turbines in the offshore wind farm within the region is expressed as , where P represents the number of wind turbines, M represents the longitude, latitude, and altitude of the wind turbine; represents the set of geographical information of the wind turbines;

[0101] The process of constructing the two-dimensional static plane features is as follows:

[0102] S11: Calculate the normalized distance ND between the wind turbine i and the wind turbine j . The calculation process is as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108] Among them, represents the longitude of the wind turbine i , represents the latitude of the wind turbine i , represents the altitude of the wind turbine i , hav () is an intermediate operation function. For any x , , R represents the radius of the earth, represents the distance between the wind turbine i and the wind turbine j , ;

[0109] S12: Represent the wind power i of the wind turbine as I , represent the wind power j of the wind turbine as J , and calculate the normalized mutual information NMI between the wind turbine i and the wind turbine j . The expression is:

[0110]

[0111]

[0112]

[0113]

[0114] Among them, is the wind turbine i and the wind turbine j of the joint distribution, is the marginal distribution of the wind turbine i ; represents the information entropy of the wind turbine i ; represents the information entropy of the wind turbine j ; represents the information entropy of the wind turbine i and the wind turbine j of the joint information entropy, represents the mutual information of the wind turbine i and the wind turbine j ;

[0115] S13: Construct a two-dimensional static plane feature based on the normalized distance and normalized mutual information between wind turbine pairs . Specifically, with the normalized distance as the horizontal axis and the normalized mutual information as the vertical axis, construct a two-dimensional static plane feature space, plot the values of the wind turbine pairs on these two feature dimensions in the feature space to form a feature map. Each wind turbine pair corresponds to a point on the feature map, and its horizontal and vertical coordinates are the normalized distance and normalized mutual information of the wind turbine pair, respectively.

[0116] In this embodiment, meteorological data in the acquisition area is collected. The meteorological data includes: wind speed V , wind direction D , temperature T , humidity H and air pressure P a . The meteorological data changes with time. The meteorological data is preprocessed. In this embodiment, the data is initially processed by min - max normalization.

[0117] The altitude of the offshore wind farm in the area L is divided into n levels, denoted as: , the i th level L i has a fixed altitude z interval, satisfying: , where represents the lower limit of the altitude interval of the i th level L i , Indicates the i th level L i upper limit of the altitude range;

[0118] At each altitude z, the meteorological data characteristics M ( z ) are represented as a multi-dimensional vector, and the expression is:

[0119] ;

[0120] The i th level L i meteorological data characteristics M ( L i ) are represented as:

[0121] .

[0122] In this embodiment, the process of dimensionality reduction for the multi-dimensional time-series wind-electricity meteorological characteristics at each altitude of each level is as follows:

[0123] S21: Set a fixed sliding window Q, and select Q consecutive sampling points from the time series to form a subsequence , where t is the current time point, and x represents any one of the meteorological variables, and the meteorological variables include wind speed V , wind direction D , temperature T , humidity H and air pressure P a ;

[0124] S22: On each subsequence q , construct a time series feature vector for each meteorological variable separately:

[0125]

[0126] Among them, each row represents the time series characteristics of a single meteorological variable of a single node within each altitude of each level, represents the time series mean of the collection point q, represents the time series variance of the collection point q, represents the time autocorrelation coefficient of the collection point q, represents the spatial autocorrelation of the collection point q; all meteorological feature vectors of the node u within each altitude of each level are:

[0127]

[0128] Integrate all node features to obtain all node feature matrices within each altitude level. f :

[0129]

[0130] S23: For the feature matrix f Perform principal component analysis to obtain a new feature matrix after dimensionality reduction , for Calculate the covariance matrix C and perform eigenvalue decomposition on it to obtain eigenvalues and corresponding eigenvectors and , select the eigenvectors corresponding to the top k largest eigenvalues to form a projection matrix W, and obtain the feature matrix after dimensionality reduction within each altitude level The expression of is:

[0131]

[0132] Among them, The size of is n * k , 2 < k < m, n is the number of meteorological variables, m is the total number of eigenvalues after decomposition.

[0133] , , and The calculation expressions of are respectively:

[0134]

[0135]

[0136]

[0137]

[0138] Among them, is the spatial weight between the position of the fan i and the position of the fan j .

[0139] The process of step S3 is as follows:

[0140] S31: Adopt the method of dynamic feature selection and set the weight allocation strategy so that the weight of the feature matrix in the high altitude layer is larger and the sum of all weights is 1; the weight allocation strategy for a single altitude layer is:

[0141]

[0142] Among them, z is the height of the L i th altitude layer, is an adjustable parameter used to control the sensitivity of the weight to the altitude of different levels, achieving the control of the differential sensitivity to high and low altitudes, is the growth index used to amplify the weight of the high-altitude layer, is the smoothing regulation coefficient;

[0143] Through weighted summation, the feature matrices after dimensionality reduction of the altitude of each level are merged into a local feature matrix

[0144]

[0145] Among them, N is the number of single altitude layers, is the th weight of the altitude layer;

[0146] S32: Use a multi-layer perceptron MLP to expand the dimension of the two-dimensional static plane feature so that the dimension of the two-dimensional static plane feature is the same as the dimension of the local feature matrix The two-dimensional static plane feature matrix after dimension expansion and the local feature matrix are vertically concatenated into a global spatial graph feature F .

[0147] In this embodiment, the multi-layer perceptron MLP includes an input layer, a hidden layer, and an output layer. When designing the multi-layer perceptron MLP, the number of nodes in the input layer matches the number of two-dimensional static plane features, that is, 2 nodes. In this embodiment, one or more hidden layers are designed, and each hidden layer contains a certain number of nodes. The number of nodes can be adjusted according to the complexity of the task and the dimension of the required output. The hidden layer uses an activation function (such as ReLU, Sigmoid, or Tanh) to introduce non-linearity, enabling the MLP to approximate complex non-linear relationships.

[0148] The process of using a multi-layer perceptron MLP to expand the dimension of the two-dimensional static plane feature is as follows:

[0149] Preprocess the two-dimensional static plane feature ;

[0150] Input the preprocessed two-dimensional static plane feature into the trained multi-layer perceptron MLP;

[0151] Calculate the one-dimensional vector output by the output layer of the multi-layer perceptron through the forward propagation algorithm of the multi-layer perceptron MLP;

[0152] Convert the one-dimensional vector output by the output layer into a matrix form with the same dimension as the local feature matrix of.

[0153] In this embodiment, the process of generating the final latent feature vector described in step S4 is as follows:

[0154] S41: Construct the adjacency matrix A F according to the cosine similarity between the wind turbine nodes in the global spatial graph feature h , and the expression is:

[0155] ;

[0156] S42: Perform forward convolution operations on the global spatial graph feature F and the adjacency matrix A h in the graph convolutional neural network model with several layers; among them, the convolution operation of any layer in the graph convolutional neural network model satisfies:

[0157]

[0158] where E is the degree matrix, is the weight matrix of the first layer of the graph convolutional neural network model layer, is the feature matrix after the first layer of graph convolution operation, is the p layer graph convolution weight matrix; in this embodiment, use PyTorch and PyTorch Geometric libraries to implement the convolution operation of the graph convolutional neural network model. First, define the graph data, including the global spatial graph feature and the adjacency matrix, use the torch_geometric.nn.conv.GCNConv class in PyTorch Geometric to define the graph convolutional neural network model layer, then stack the graph convolutional neural network model layers to create a multi-layer graph convolutional neural network model, and then perform forward propagation. After performing multi-layer graph convolution operations, extract the output features of the last layer of the graph convolutional neural network model, and the output features include the final latent feature vector of the wind turbines in the offshore wind farm.

[0159] The process of step S5 is:

[0160] Use the latent feature vector and the corresponding historical wind power data as training data to train the linear regression prediction model through forward propagation;

[0161] Calculate the prediction error using the loss function, update the model parameter weights of the linear regression prediction model using the optimizer, and end the training after the loss function converges or reaches the predetermined number of training rounds to obtain a trained linear regression prediction model;

[0162] Execute steps S1 to S4 to obtain a new latent feature vector, input the new latent feature vector into the trained linear regression prediction model, and output the wind power of the offshore wind farm.

[0163] Embodiment 2

[0164] In this embodiment, to verify the effectiveness of the proposed offshore wind power prediction method with multi-level spatio-temporal perception, in step S1, obtain the geographical location information of the wind turbines in an offshore wind farm and preprocess the data;

[0165] In step S2, obtain the measured meteorological data of an offshore wind farm, including the power of the offshore wind farm, the wind speed, wind direction, temperature, humidity, and air pressure at each altitude, and preprocess the data. Divide the altitude into 8 levels: 0m, 50m, 100m, 150m, 200m, 250m, 300m, and 350m;

[0166] Use the proposed offshore wind power prediction method with multi-level spatio-temporal perception for prediction, and obtain the prediction effect diagram of the offshore wind power as shown in Figure 2 In Figure 2 , the dotted trend line represents the predicted value of the offshore wind power, and the solid trend line represents the true value of the offshore wind power. It can be seen from the two curves shown in Figure 2 that the error between the true value of the wind power and the predicted value of the wind power is very small, indicating that the method proposed by the present invention can effectively improve the accuracy of the offshore wind power prediction.

[0167] Embodiment 3

[0168] As shown in Figure 3 , the present application proposes an offshore wind power prediction system with multi-level spatio-temporal perception, which is characterized in that the system is used to implement the proposed offshore wind power prediction method with multi-level spatio-temporal perception, and includes:

[0169] A two-dimensional static plane feature module, which regards each wind turbine in the offshore wind farm within the region as a node, determines the spatial relationship of the wind turbines according to the geographical location information of each wind turbine and the distance between the wind turbines, and constructs two-dimensional static plane features;

[0170] A dimensionality reduction module, which obtains multi-dimensional time-series wind power meteorological features at multi-level altitudes and reduces the dimensionality of the multi-dimensional time-series wind power meteorological features at each level of altitude;

[0171] A feature integration and merging module is used to assign weights to the multi-dimensional time-series wind power and meteorological features after dimensionality reduction, obtain the feature matrix after dimensionality reduction for each hierarchical altitude, construct local graph features based on the feature matrix after dimensionality reduction, and merge the local graph features and two-dimensional static plane features into a global spatial graph feature;

[0172] An extraction module inputs the global spatial graph feature into a graph convolutional neural network model to generate a final latent feature vector;

[0173] A prediction module uses the final latent feature vector to train a pre-constructed linear regression prediction model, and uses the trained linear regression prediction model for wind power prediction of an offshore wind farm.

[0174] The embodiments are only examples for clearly illustrating the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for offshore wind power prediction with multi-level spatiotemporal perception, characterized in that: The following steps are involved: S1: Consider each wind turbine in the offshore wind farm in the region as a node, determine the spatial relationship of the wind turbines based on the geographical location information of each wind turbine and the distance between the wind turbines, and construct a two-dimensional static plane feature; S2: Obtain multi-dimensional time series wind power meteorological characteristics at multiple levels of altitude, and reduce the dimension of the multi-dimensional time series wind power meteorological characteristics at each level of altitude; The process of reducing the dimension of multi-dimensional time series wind power meteorological characteristics at each level of altitude is as follows: S21: Set a fixed sliding window Q and select Q consecutive sampling points from the time series to form a subsequence ,in, t is the current time point, x represents any one of the meteorological variables, including wind speed V ,wind direction D ,temperature T ,humidity H and air pressure P a ; S22: In each subsequence q A time series feature vector is constructed for each meteorological variable. : Each row represents the time series characteristics of a single meteorological variable at a single node at each level of altitude. represents the time series mean of the acquisition point q, represents the time series variance of the acquisition point q, represents the time autocorrelation coefficient of the acquisition point q, Represents the spatial autocorrelation of the collection point q; the nodes within each level of altitude u All meteorological feature vectors are: Integrate all node features to obtain the feature matrix of all nodes at each level of altitude f : S23: Feature matrix f Perform principal component analysis to obtain a new feature matrix after dimensionality reduction ,right Calculate the covariance matrix C and perform eigenvalue decomposition on it to obtain the eigenvalues ​​and corresponding eigenvectors and , select before k The eigenvectors corresponding to the largest eigenvalues ​​form the projection matrix W , and obtain the feature matrix after the dimension reduction of each level altitude The expression is: in, The size is n * k , 2<k<m, n is the number of meteorological variables, m is the total number of eigenvalues ​​after eigendecomposition; S3: Assign weights to the multi-dimensional time-series wind power meteorological features after dimensionality reduction, obtain the feature matrix after dimensionality reduction at each level of altitude, construct local graph features based on the feature matrix after dimensionality reduction, and merge the local graph features and the two-dimensional static plane features into a global space graph feature; S4: Input the global spatial graph features into the graph convolutional neural network model to generate the final latent feature vector; S5: The final potential feature vector is used to train the constructed linear regression prediction model, and the trained linear regression prediction model is used for wind power prediction of offshore wind farms.

2. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 1 is characterized in that: The geographical location information of each wind turbine includes: the longitude, latitude and altitude of each wind turbine; the geographical information of the wind turbines in the offshore wind farm in the region is represented as ,in P Indicates the number of fans, M Indicates the longitude, latitude and altitude of the wind turbine; A collection representing geographic information of wind turbines; The process of constructing a 2D static plane feature is: S11: Calculate wind turbines in offshore wind farms in the region i With fan j The normalized distance ND between them is calculated as follows: in, Indicates fan i Longitude, Indicates fan i The latitude, Indicates fan i The altitude, have () is the intermediate operation function. For any x , , R represents the radius of the earth, Indicates fan i With fan j The distance between ; S12: The fan i Wind power Expressed as I , fan j Wind power Expressed as J , calculate the fan i and fan j The normalized mutual information NMI between is expressed as: in, It is a fan i and fan j The joint distribution of It is a fan i The marginal distribution of Indicates fan i The information entropy of Indicates fan j The information entropy of Indicates fan i and fan j The joint information entropy of Indicates fan i and fan j The mutual information of S13: Constructing 2D static plane features based on normalized distance and normalized mutual information between wind turbine pairs .

3. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 2 is characterized in that: Collect meteorological data in the area, including: wind speed V ,wind direction D ,temperature T ,humidity H and air pressure P a , the meteorological data changes with time, and the meteorological data is preprocessed; The altitude of offshore wind farms in the region L Divide into n levels, expressed as: , No. i Levels L i The interval of altitude z is fixed and satisfies: ,in, Indicates i Levels L i The lower limit of the altitude range is Indicates i Levels L i The upper limit of the altitude range; At each altitude z, the meteorological data features M ( z ) is expressed as a multidimensional vector, and the expression is: ； No. i Levels L i Meteorological data characteristics M ( L i ) is expressed as: 。 4. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 1 is characterized in that: , , and The calculation expressions are: in, For fans i Location and fan j The spatial weights between the locations.

5. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 4 is characterized in that: The process of step S3 is: S31: Using dynamic feature selection, set the weight allocation strategy so that the feature matrix weight of the high altitude layer is larger and the sum of all weights is 1; the weight allocation strategy of the single altitude layer is: in, z For the L i The height of the altitude layer, It is an adjustable parameter used to control the sensitivity of weights to different levels of altitude, and to achieve differential sensitivity control for high altitude and low altitude. is a growth index used to amplify the weight of high altitude layers. is the smoothing control coefficient; Through weighted summation, the feature matrix after all the altitudes of each level are reduced in dimension Merge into a local feature matrix : in, N is the number of single altitude layers, For the The weight of each altitude layer; S32: Use multi-layer perceptron MLP to transform two-dimensional static plane features The dimension expansion makes the two-dimensional static plane feature The dimension and local feature matrix The dimension of the two-dimensional static plane feature matrix after the dimension expansion is the same as and the local feature matrix Vertical splicing into global spatial graph features F .

6. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 5 is characterized in that: Multi-layer perceptron MLP is used to transform the two-dimensional static plane features The process of dimensional expansion is: For 2D static plane features Perform pre-processing; The preprocessed two-dimensional static plane features Input to the trained multi-layer perceptron MLP; Calculate the one-dimensional vector output by the multi-layer perceptron output layer through the forward propagation algorithm of the multi-layer perceptron MLP; Convert the one-dimensional vector of the output layer into the local feature matrix Matrix form of the same dimensions.

7. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 6 is characterized in that: The process of generating the final potential feature vector described in step S4 is: S41: Based on the global spatial graph features F Cosine similarity between wind machine nodes, constructing adjacency matrix A h , the expression is: ； S42: In a graph convolutional neural network model with several layers, global spatial graph features F and the adjacency matrix A h Perform a forward convolution operation; wherein the convolution operation of any layer in the graph convolutional neural network model satisfies: Where E is the degree matrix, is the weight matrix of the first graph convolutional neural network model layer, is the feature matrix after the first layer of graph convolution operation, It is p A weight matrix of a layer graph convolution; after performing a multi-layer graph convolution operation, extracting the output features of the last layer of the graph convolution neural network model, wherein the output features include the final potential feature vector of the offshore wind farm wind turbine.

8. The offshore wind power prediction method with multi-level spatiotemporal perception according to claim 7 is characterized in that: The process of step S5 is: The potential feature vector and the corresponding historical wind power data are used as training data to train the linear regression prediction model through forward propagation; The prediction error is calculated using the loss function, and the model parameter weights of the linear regression prediction model are updated using the optimizer. After the loss function converges or reaches the predetermined number of training rounds, the training is terminated to obtain a trained linear regression prediction model. Execute steps S1 to S4 to obtain a new potential feature vector, input the new potential feature vector into the trained linear regression prediction model, and output the wind power of the offshore wind farm.

9. An offshore wind power prediction system with multi-level spatiotemporal perception, characterized in that: The system is used to implement the offshore wind power prediction method with multi-level time and space perception as described in any one of claims 1 to 8, comprising: The two-dimensional static plane feature module is used to regard each wind turbine in the offshore wind farm in the region as a node, determine the spatial relationship of the wind turbines according to the geographical location information of each wind turbine and the distance between the wind turbines, and construct the two-dimensional static plane feature; A dimension reduction module is used to obtain multi-dimensional time series wind power meteorological characteristics at multiple levels of altitude, and to reduce the dimension of the multi-dimensional time series wind power meteorological characteristics at each level of altitude; The feature integration and merging module is used to assign weights to the multi-dimensional time-series wind power meteorological features after dimensionality reduction, obtain the feature matrix after dimensionality reduction at each level of altitude, construct local graph features based on the feature matrix after dimensionality reduction, and merge the local graph features and the two-dimensional static plane features into a global spatial graph feature; The extraction module inputs the global spatial graph features into the graph convolutional neural network model to generate the final latent feature vector; The prediction module uses the final potential feature vector to train the constructed linear regression prediction model, and uses the trained linear regression prediction model for wind power prediction of offshore wind farms.

Images