Wind speed prediction method and device and storage medium

Abstract

The invention discloses a wind speed prediction method and device and a storage medium, and belongs to the technical field of wind speed prediction.The method comprises the steps that a multivariable meteorological time sequence collected by a plurality of meteorological stations in a target area is obtained and then preprocessed, and the multivariable meteorological time sequence comprises the values of a plurality of meteorological elements arranged according to the time sequence; inputting the preprocessed multivariable meteorological time sequence into the trained wind speed prediction model to obtain a wind speed prediction result of a future set time step; wherein a multi-scale decomposable hybrid module based on a frequency domain multilayer perceptron, a multilayer graph convolutional network based on a hypervariable graph and a space graph and an Informer prediction module based on inverted text embedding are introduced into the wind speed prediction model. According to the method, extraction of multivariable attributes, complex interaction relations and multi-time scale features is achieved by introducing the modules, model lightweight is achieved, and the problems of low accuracy and low efficiency in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to a wind speed prediction method, device and storage medium, belonging to the wind speed prediction technical field. BACKGROUND

[0002] Wind speed prediction is a key link in many fields such as modern weather forecasting, renewable energy scheduling (such as wind power), aviation and shipping safety, and has important research value and engineering significance. Traditional wind speed prediction methods mainly include numerical weather prediction models, statistical regression models, and machine learning models based on single station time series. However, these methods often face problems such as limited modeling ability, insufficient spatial relationship description, sensitivity to non-stationary and high noise data, and are difficult to meet the demand of high-precision wind speed prediction.

[0003] In recent years, with the rapid development of deep learning technology, time series prediction methods based on neural networks have gradually become a research hotspot. Among them, the Transformer (a neural network architecture based on attention mechanism) is widely used in meteorological element prediction tasks due to its strong global modeling ability. However, the standard Transformer model has problems such as large computational overhead when processing long time series and insufficient ability to model spatial structure. At the same time, traditional spatial modeling methods often construct fixed topological structures based on meteorological stations, ignoring the multivariate attributes of meteorological data and their complex interaction relationships, resulting in weak learning ability of the prediction model for the coupling relationship between variables.

[0004] In addition, wind speed data usually has obvious multi-time scale variation characteristics and high frequency noise interference. Most existing models use fixed scale input, which is difficult to fully extract the variation patterns at different time scales, affecting the prediction accuracy and generalization ability of the model.

[0005] In summary, the existing technology has the problems of low accuracy and low efficiency in wind speed prediction. SUMMARY

[0006] The purpose of the present application is to provide a wind speed prediction method, device and storage medium, which realizes the extraction of multivariate attributes and their complex interaction relationships and multi-time scale characteristics, and realizes the lightweight of the model, to solve the problems of low accuracy and low efficiency in the wind speed prediction process of the existing technology.

[0007] To achieve the above purpose, the present application is realized by adopting the following technical solutions:

[0008] In a first aspect, the present application provides a wind speed prediction method, comprising:

[0009] Obtaining a plurality of multivariate weather time series collected by a plurality of weather stations in a target area, the plurality of multivariate weather time series comprising values of a plurality of weather elements arranged in time sequence;

[0010] Inputting the preprocessed multivariate weather time series into a trained wind speed prediction model;

[0011] Performing multi-scale enhancement on the preprocessed multivariate weather time series by a multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence;

[0012] Inputting the multi-scale enhanced sequence into a multi-layer graph convolution network of the wind speed prediction model, taking values of weather variables as node features, determining adjacency relationships between nodes according to feature similarities between nodes, extracting variable interaction features in the multi-scale enhanced sequence through frequency domain graph convolution in a hyper-variable graph, taking weather stations as nodes, determining adjacency relationships between nodes according to geographical distance correlations between nodes, extracting spatiotemporal features in the multi-scale enhanced sequence through residual graph convolution in a spatial graph;

[0013] Performing inverted text embedding and low-sparsity attention operation on the spatiotemporal features by a prediction module of the wind speed prediction model to obtain an attention matrix, inputting the attention matrix into layer normalization and a feedforward neural network to obtain variable fusion high-dimensional features;

[0014] Converting the variable fusion high-dimensional features into wind speed prediction results by a fully connected layer of the wind speed prediction model.

[0015] Further, the weather elements include temperature, air pressure, wind speed, wind direction, dew point, and rainfall.

[0016] Further, performing multi-scale enhancement on the preprocessed multivariate weather time series by the multi-scale decomposable hybrid module of the wind speed prediction model to obtain the multi-scale enhanced sequence includes: performing hierarchical downsampling and concatenation on the preprocessed multivariate weather time series to obtain a downsampled and concatenated matrix;

[0017] Processing the downsampled and concatenated matrix through a frequency domain multi-layer perceptron to obtain spectral features at different time scales;

[0018] Fusing the spectral features at different time scales to obtain the multi-scale enhanced sequence.

[0019] Further, the hierarchical downsampling is completed by hierarchical average pooling;

[0020] Processing the downsampled and concatenated matrix through a frequency domain multi-layer perceptron to obtain spectral features at different time scales includes performing the following steps of extracting spectral features at a single time scale in all time scales:

[0021] transforming the down-sampling concatenation matrix in a single time scale to frequency domain;

[0022] processing the real part and the imaginary part of the down-sampling concatenation matrix transformed to frequency domain respectively through frequency domain multi-layer perception, and then restoring to time domain to obtain the spectral feature in a single time scale.

[0023] Further, the variable interaction features in the multi-scale enhanced sequence are extracted through frequency domain graph convolution in the hyper-variable graph, and are processed through the following formula:

[0024] ;

[0025] wherein, represents the multi-scale enhanced sequence, represents the discrete Fourier transform, represents the inverse Fourier transform, represents the transformation matrix in the frequency domain set, represents the variable interaction feature.

[0026] Further, in the hyper-variable graph, the expression of the adjacency matrix between any two nodes is:

[0027] ;

[0028] wherein, represents the adjacency matrix between two nodes in the hyper-variable graph, represents the natural exponential function, represents the value of the meteorological variable represented by the i-th node in the hyper-variable graph, represents the value of the meteorological variable represented by the j-th node in the hyper-variable graph, represents the value of the meteorological variable represented by the i-th node in the hyper-variable graph, represents the value of the meteorological variable represented by the j-th node in the hyper-variable graph, represents the scale parameter set, represents the connection threshold value set between meteorological variables, represents the norm.

[0029] Further, the space-time features in the multi-scale enhanced sequence are extracted through residual graph convolution in the space graph, and are processed through the following formula:

[0030] ;

[0031] wherein, represents the variable interaction feature; represents the adjacency matrix between two nodes in the space graph, which is constructed according to the geographical positions of the meteorological stations corresponding to the two nodes; represents the residual graph convolution network; The spatio-temporal features represent local variable dependencies and global spatial topology information in a multi-layer nested graph structure.

[0032] Further, the spatio-temporal features are inverted text embedding and low-rank sparse attention operation to obtain an attention matrix:

[0033] The complete spatio-temporal sequence of a meteorological variable in all time steps and all cities in the spatio-temporal features is taken as a word element, and all word elements are combined into a matrix in the order of features and then transposed to obtain an inverted embedding sequence;

[0034] The inverted embedding sequence is respectively subjected to different linear transformations to obtain a query matrix and a key matrix;

[0035] The query matrix and the key matrix are respectively multiplied by the corresponding low-rank projection matrix to obtain the low-rank projected query matrix and the low-rank projected key matrix;

[0036] According to the low-rank projected query matrix and the low-rank projected key matrix,

[0037] The attention matrix is calculated by the following formula:

[0038] ;

[0039] Among them, the attention matrix, the normalized exponential function, the low-rank projected query matrix, the low-rank projected key matrix, a scaling factor used to prevent too large or too small in calculating the inner product, transpose.

[0040] In a second aspect, the present application provides a wind speed prediction device, comprising:

[0041] A multivariate acquisition module is configured to: after acquiring a multivariate meteorological time sequence collected by a plurality of meteorological stations in a target area, the multivariate meteorological time sequence includes values of a plurality of meteorological elements arranged in time sequence;

[0042] A wind speed prediction module is configured to: input the preprocessed multivariate meteorological time sequence into a trained wind speed prediction model;

[0043] A multi-scale decomposable hybrid module of the wind speed prediction model is configured to perform multi-scale enhancement on the preprocessed multivariate meteorological time sequence to obtain a multi-scale enhanced sequence;

[0044] input the multi-scale enhanced sequence to a multi-layer graph convolution network of the wind speed prediction model, represent nodes by values of meteorological variables, determine adjacency relationships between nodes by feature similarity between nodes, extract variable interaction features in the multi-scale enhanced sequence by frequency domain graph convolution in a hyper-variable graph, and represent nodes by meteorological stations, determine adjacency relationships between nodes by geographical distance correlation between nodes, extract spatio-temporal features in the multi-scale enhanced sequence by residual graph convolution in a spatial graph;

[0045] perform inverted text embedding and low-rank sparse attention operation on the spatio-temporal features by a prediction module of the wind speed prediction model to obtain an attention matrix, input the attention matrix to layer normalization and a feedforward neural network to obtain variable fusion high-dimensional features;

[0046] convert the variable fusion high-dimensional features into a wind speed prediction result by a fully connected layer of the wind speed prediction model.

[0047] In a third aspect, the present application provides a computer readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the wind speed prediction method of any one of the first aspect.

[0048] Compared with the prior art, the present application has the following beneficial effects:

[0049] The wind speed prediction method, device and storage medium provided by the present application significantly enhance the modeling capability of the model for different frequency components in the non-stationary wind speed time series by introducing a multi-scale decomposable hybrid module based on a frequency domain multi-layer perceptron, and suppress the influence of high-frequency noise on the prediction result. At the same time, a hyper-variable graph is constructed and combined with a graph convolution network to model at both variable and spatial levels, thereby efficiently extracting complex dependency relationships in multi-site and multi-variable wind speed data. In the prediction stage, an improved Informer structure is introduced to significantly improve the ability to capture long-time dependency information by using an inverted Embedding mechanism and a low-rank sparse attention mechanism, and to reduce the computational complexity of the model. Compared with existing wind speed prediction methods, the present application has significant advantages in accuracy, stability and computational efficiency, and has good generalization ability and practicality in multi-variable interaction modeling and spatio-temporal fusion prediction tasks.

[0050] The present application is suitable for wind farm ultra-short-term wind speed prediction, power system dynamic scheduling and meteorological warning application scenarios, and has important engineering popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is a flowchart of a wind speed prediction method corresponding to example 1;

[0052] Figure 2 is a schematic diagram of the technical concept provided by the present application;

[0053] Figure 3 is a flowchart of a wind speed prediction method corresponding to embodiment 2;

[0054] Figure 4 is a schematic diagram of the processing flow of the multi-scale decomposable hybrid module provided by the present application on its input. DETAILED DESCRIPTION

[0055] The present application will be further described below in conjunction with the accompanying drawings, and the following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0056] Embodiment 1

[0057] As shown in the figure, the present embodiment provides a wind speed prediction method, comprising: Figure 1

[0058] obtaining and preprocessing the multivariate weather time series collected by multiple weather stations in the target area, wherein the multivariate weather time series includes values of multiple weather elements arranged in time sequence;

[0059] inputting the preprocessed multivariate weather time series into the trained wind speed prediction model;

[0060] performing multi-scale enhancement on the preprocessed multivariate weather time series through the multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence;

[0061] inputting the multi-scale enhanced sequence into the multi-layer graph convolution network of the wind speed prediction model, taking the values of the weather variables as nodes and determining the adjacency relationship between the nodes according to the feature similarity between the nodes, extracting the variable interaction features in the multi-scale enhanced sequence through frequency domain graph convolution in the hyper-variable graph; taking the weather stations as nodes and determining the adjacency relationship between the nodes according to the geographical distance correlation between the nodes, extracting the spatiotemporal features in the multi-scale enhanced sequence through residual graph convolution in the spatial graph;

[0062] performing inverted text embedding and low-rank sparse attention operation on the spatiotemporal features through the prediction module of the wind speed prediction model to obtain an attention matrix, inputting the attention matrix into the layer normalization and feedforward neural network to obtain variable fusion high-dimensional features;

[0063] converting the variable fusion high-dimensional features into wind speed prediction results through the fully connected layer of the wind speed prediction model.

[0064] ​The application significantly enhances the modeling ability of the model for different frequency components in the non-stationary wind speed time series by introducing a multi-scale decomposable hybrid module based on a frequency domain multi-layer perception, and suppresses the influence of high-frequency noise on the prediction result. At the same time, an ultra-variable graph is constructed and combined with a graph convolution network to model at the variable level and the spatial level, thereby realizing efficient extraction of complex dependency relationships in multi-site and multi-variable wind speed data. In the prediction stage, an improved Informer structure is introduced to significantly improve the ability to capture long-time dependency information and reduce the computational complexity of the model. Compared with existing wind speed prediction methods, the application has significant advantages in accuracy, stability and computational efficiency, and has good generalization ability and practicability in multi-variable interactive modeling and spatio-temporal fusion prediction tasks.

[0065] Embodiment 2

[0066] According to Figure 2 The technical concept shown in the flowchart shown in Figure 3 The present embodiment provides a wind speed prediction method, comprising the following first to sixth steps.

[0067] Step 1, data set normalization.

[0068] The data set used by the application is from two actual meteorological data sets of region one and region two. The actual meteorological data set from region two contains four meteorological elements (temperature, air pressure, wind speed and wind direction) from five cities, recorded once an hour. The actual meteorological data set from region one contains meteorological data from seven cities, including wind speed, wind direction, temperature, dew point, air pressure and rainfall. The shape of the two actual meteorological data sets is , wherein, represents the number of meteorological elements, is the number of time steps, is the number of cities. The data in the two actual meteorological data sets are arranged in time sequence as multi-variable meteorological time series respectively.

[0069] In order to perform data preprocessing, maximum and minimum normalization is applied to each data set to scale all data to the range of [0, 1]. The normalization formula is as follows:

[0070] (1);

[0071] , wherein, represents the original data in the actual meteorological data set, and are the minimum and maximum values in the original data respectively, is the normalized data.

[0072] The second step is to extract multi-scale features from the multi-scale decomposable hybrid module in the wind speed prediction model.

[0073] like Figure 4 As shown in the figure, in order to effectively model the multi-time scale features in wind speed data, the present invention designs a multi-scale decomposable hybrid module based on a frequency domain multi-layer perceptron, which includes the following three parts:

[0074] (1) Multi-scale downsampling decomposition:

[0075] The input multivariate meteorological time series is downsampled by layered average pooling to obtain a multi-time scale representation. Assume that the input of the multi-scale decomposable mixing module is , Represents a shape For a dataset, let the maximum number of downsampling layers be , then the multiple time scales are expressed as:

[0076] (2);

[0077] in, Indicates the Layer average pooling, Indicates the Multi-timescale representation of layer average pooling output.

[0078] For multiple sequences obtained by downsampling, splice them according to the channel dimension to obtain the downsampling splicing matrix , and the downsampled splicing matrix is ​​used as input to the next step.

[0079] (2) Frequency Domain Multilayer Perceptron (FMLP):

[0080] The downsampled concatenated matrix at each time scale is converted to the frequency domain through a discrete Fourier transform (DFT):

[0081] (3);

[0082] in, represents the down-sampling concatenation matrix converted to the frequency domain, represents the discrete Fourier transform, is the downsampling stitching matrix.

[0083] Then, in the frequency domain, the real and imaginary parts of the downsampled concatenated matrix converted to the frequency domain are processed using a multilayer perceptron:

[0084] (4);

[0085] (5);

[0086] wherein, represents a real part of the down-sampling concatenation matrix converted to the frequency domain, represents an imaginary part of the down-sampling concatenation matrix converted to the frequency domain, and is a weight matrix, and is a bias term, is an activation function, represents a result of using a multi-layer perceptron to process the real part of the down-sampling concatenation matrix converted to the frequency domain, represents a result of using a multi-layer perceptron to process the imaginary part of the down-sampling concatenation matrix converted to the frequency domain.

[0087] is composed of the down-sampling concatenation matrix processed by the multi-layer perceptron and is then recovered to the time domain using an inverse Fourier transform:

[0088] (6);

[0089] wherein, represents the down-sampling concatenation matrix processed by the multi-layer perceptron, represents an inverse Fourier transform, represents a spectral feature in a single time scale in the time domain.

[0090] (3) Residual fusion

[0091] In order to better fuse the spectral features of different time scales and avoid loss of features, the application adopts the following residual fusion mechanism:

[0092] (7);

[0093] wherein, represents a down-sampling concatenation matrix, represents a spectral feature in a single time scale in the time domain, represents a discrete Fourier transform, represents a multi-scale enhanced sequence, and the multi-scale enhanced sequence is an input of the multi-layer graph convolutional network.

[0094] The third step is to extract the space-time features in the multi-layer graph convolutional network in the wind speed prediction module.

[0095] ​​In the wind speed prediction task, in order to effectively capture the complex spatial dependence relationship between different meteorological stations and the multivariate inside the same meteorological station, the present application proposes a multi-layer graph convolution network, including two levels of hyper-variable graph modeling and spatial graph convolution modeling between stations, and realizes efficient spatio-temporal modeling capability through nested fusion.

[0096] The traditional modeling method taking meteorological stations as graph nodes can only represent the spatial relationship between stations, and it is difficult to depict the interaction between different meteorological variables in the same station. Therefore, the present application introduces a hyper-variable graph network (HGN), which regards multiple meteorological variables (such as wind speed, temperature, humidity, etc.) in each meteorological station as independent graph nodes, establishes a graph structure between the variables, and enhances the expression ability of modeling.

[0097] In the hyper-variable graph, the adjacency relationship between any two nodes is controlled by the feature similarity, and the adjacency matrix A is defined as follows:

[0098] (8);

[0099] Wherein, A represents the adjacency matrix between two nodes in the hyper-variable graph, e represents the natural exponential function, x represents the value of the meteorological variable represented by the i th node in the hyper-variable graph, x represents the value of the meteorological variable represented by the i th node in the hyper-variable graph, represents the set scale parameter, represents the set connection threshold between meteorological variables, represents the norm. After establishing the hyper-variable graph, the frequency domain graph convolution method is used to extract the variable interaction features in the multi-scale enhanced sequence (the output features of the fused variable dependence after frequency domain graph convolution), so as to reduce the calculation complexity and improve the modeling efficiency:

[0100]

[0101] (9);

[0102] Wherein, X represents the multi-scale enhanced sequence, DFT represents the discrete Fourier transform, IDFT represents the inverse Fourier transform, F represents the transformation matrix in the frequency domain, represents the variable interaction feature.

[0103] ​​​After completing the hyper-variable graph modeling inside the weather station, the application further constructs the spatial graph structure between the weather stations, and uses residual graph convolution (Residual Graph Convolutional Network, RGCN) for modeling, realizing the information fusion from the local (inside the weather station) to the global (between the weather stations).

[0104] First, the output of the aforementioned hyper-variable graph is taken as the input of the spatial graph modeling module. As the node feature, the node of the spatial graph is defined as the weather station, and the connection relationship between the nodes is represented by the adjacency matrix A. Residual graph convolution operation is used in the spatial graph, and the process is as follows:

[0105] The spatial graph structure is established with the weather station as the node, and the residual graph convolution is performed:

[0106] (10);

[0107] wherein, x represents the output of the hyper-variable graph network, x represents the output of the hyper-variable graph network, A represents the adjacency matrix between two nodes in the spatial graph, and the adjacency matrix between two nodes in the spatial graph is constructed according to the geographical positions of the weather stations corresponding to the two nodes, RGCN represents the residual graph convolution network, x represents the spatiotemporal feature containing the local variable dependence and global spatial topological information in the multi-layer nested graph structure, and the spatiotemporal feature x is taken as the input of the prediction module. x represents the spatiotemporal feature containing the local variable dependence and global spatial topological information in the multi-layer nested graph structure, and the spatiotemporal feature x is taken as the input of the prediction module.

[0108] In the fourth step, the Informer prediction module based on the inverted Embedding predicts the wind speed at the future preset time step. The Informer is a variant of the Transformer.

[0109] After completing the multi-scale time feature extraction and multi-layer graph convolution modeling, the application further designs an efficient time series prediction mechanism, that is, the Informer prediction module based on the inverted Embedding, which is named iInformer, for finally generating the wind speed prediction result at the future time step. The module combines the inverted text embedding idea of iTransformer and the sparse attention mechanism of Informer, reduces the computational complexity, effectively models the long-term dependence relationship between multiple variables, and improves the prediction accuracy and efficiency.

[0110] The data organization of inverted text embedding is as follows: the traditional Transformer model treats each time step as an input word. When processing multivariate data, variables with different physical meanings need to be projected into the same embedding space, which may cause confusion between variables. At the same time, the computational complexity of this structure is very high when processing long time series. This paper adopts an inverted text embedding mechanism, treating the time series of each variable as an independent word, so that the attention mechanism can directly act between variables. The original input data is ,in Represents the characteristic dimension of each city; represents the number of time steps; Indicates the number of cities (weather stations).

[0111] Traditional Transformer uses a time-based approach to embed text, using the following formula:

[0112] (11);

[0113] in, represents the embedding sequence obtained by text embedding through traditional Transformer, represents all features of the first time step, represents all features of the second time step, Represents all features of the Nth time step. In traditional Transformer, one time step is used as a word unit, including N variables.

[0114] In the iInformer provided in this embodiment, variable-based inverted text embedding is adopted, that is, it is performed using the following formula:

[0115] (12);

[0116] in, represents the inverted embedding sequence, represents the complete sequence of the first feature at all time steps, represents the complete sequence of the second feature at all time steps, Indicates the In iInformer, the complete time series of a meteorological variable in the spatiotemporal features at all time steps is taken as a word, so that the attention mechanism can capture the long-term dependency between variables and reduce redundant calculations of time steps.

[0117] On the basis of the above inverted Embedding, the structure of the Transformer is further optimized, and a low-rank sparse attention mechanism based on a kernel function is introduced to reduce the computational burden. The attention matrix is calculated through the low-rank sparse attention mechanism based on the kernel function:

[0118] The inverted embedding sequence is subjected to different linear transformations to obtain a query matrix and a key matrix .

[0119] The query matrix and the key matrix are subjected to low-rank projection:

[0120] (13);

[0121] (14);

[0122] wherein, represents the query matrix, represents the key matrix, and represents a set low-rank projection matrix, represents the query matrix after low-rank projection, represents the key matrix after low-rank projection;

[0123] The attention matrix is calculated through the following formula:

[0124] (15);

[0125] wherein, represents the attention matrix, represents a normalized exponential function, represents a scaling factor for preventing too large or too small in the calculation of inner product, represents transposition.

[0126] In this way, when calculating attention, the iInformer reduces the computational complexity through low-rank approximation, and can effectively capture the long-distance dependence relationship between variables in multivariate sequence modeling.

[0127] The attention matrix is input into layer normalization and a feedforward neural network to obtain variable fusion high-dimensional features Finally, the variable fusion high-dimensional features are output by the Informer prediction module based on inverted text embedding.

[0128] To obtain the target prediction value, the present application introduces a fully connected layer at the end of the model to map the internal representation to the wind speed prediction space and generate the final wind speed prediction result at the future set time step. The fully connected layer can be formally represented as:

[0129] (16);

[0130] wherein, represents the wind speed prediction result, represents the ReLU activation function, and are set fully connected parameters.

[0131] Step 5, model verification and training.

[0132] After the wind speed prediction model is constructed, the wind speed prediction model provided by the present application needs to be trained and verified to ensure its effectiveness and generalization ability in the wind speed prediction task.

[0133] In order to verify the effectiveness of the model proposed in the present application in the wind speed prediction task, the present application uses two real data sets of region two and region one, uses the mean absolute error (Mean Absolute Error, MAE) and the mean squared error (Mean Squared Error, MSE) as evaluation indexes, and the definitions of MAE and MSE are as follows:

[0134] (17);

[0135] (18);

[0136] wherein, represents the real value of the i-th sample in the real data set, represents the predicted value of the i-th sample in the real data set, represents the total number of samples in the real data set. In terms of loss function, the present application selects MSE as the loss function, selects Adam (Adaptive Moment Estimation, an optimization algorithm based on gradient descent method) as the optimizer, sets the initial learning rate to 0.0001. The batch size is set to 64, the training round number is 50, and the early stopping strategy is used for the validation set.

[0137] Step 6, wind speed prediction test.

[0138] Step 6, wind speed prediction test.

[0139] ​To further verify the practicability and superiority of the wind speed prediction model (named MS-iInformer-GCNs) proposed in the application in the wind speed prediction task, two actual meteorological data sets from multiple meteorological stations in Region 2 and Region 1 are selected. In the experiment, the method provided by the application is compared with the method of using the mainstream sequence model for wind speed prediction, the mainstream sequence model includes the recurrent neural network structure (GRU), Transformer and its improved version Informer, and the wind speed prediction task is carried out under different time spans, and the prediction accuracy and stability of each model are compared.

[0140] On the actual meteorological data set from Region 2, the test task is set to predict the wind speed in the future 6 hours, 12 hours, 18 hours and 24 hours. The experimental results shown in Table 1 show that the MAE and MSE of MS-iInformer-GCNs are the lowest under all prediction lengths, among which the MAE in the tasks of predicting the future 6 hours, 12 hours, 18 hours and 24 hours is 1.155, 1.408, 1.523 and 1.624 respectively, and the MSE in the tasks of predicting the future 6 hours, 12 hours, 18 hours and 24 hours is 2.370, 3.383, 3.899 and 4.371 respectively. In comparison, the MAE of the traditional GRU in each period is higher than 1.6, and the MSE is more than 5.9 at 24h, which is significantly higher than the model provided by the application. Although Transformer and Informer have obvious improvement compared with GRU, there is still a trend of error expansion in long period prediction, especially in 24h prediction, the MSE reaches 4.978 and 4.950 respectively, while the model result of the application is 4.371, which shows that it has better stability and accuracy in long time sequence modeling.

[0141] Table 1-MAE and MSE table of wind speed prediction task on actual meteorological data set from Region 2

[0142]

[0143] On the regional-city dataset, the prediction task was set to multiple time spans ranging from 2 hours to 10 hours, and the model performance was evaluated from the perspectives of MAE and MSE. The experimental results shown in Table 1 show that MS-iInformer-GCNs still exhibits excellent prediction results. The MAEs of 2h, 4h, 6h, 8h and 10h were 7.38, 8.41, 9.35, 10.06 and 10.64, respectively, and the MSEs were 96.59, 125.10, 155.03, 180.41 and 201.76, respectively, which are superior to other models in all aspects. Especially in the long-term prediction of 10 hours, the MSE of GRU is as high as 382.54, Informer and Transformer also exceed 220, and the MSE of the model provided by the present invention is 201.76, showing a strong generalization ability for modeling complex spatiotemporal dependencies.

[0144] Table 2 - MAE and MSE of wind speed prediction task on the actual meteorological dataset from Region 1

[0145]

[0146] Example 3

[0147] This embodiment provides a wind speed prediction device, including:

[0148] The multivariate acquisition module is configured to: acquire and preprocess a multivariate meteorological time series collected by multiple meteorological stations in a target area, wherein the multivariate meteorological time series includes values ​​of multiple meteorological elements arranged in time sequence;

[0149] The wind speed prediction module is configured to: input the preprocessed multivariate meteorological time series into the trained wind speed prediction model;

[0150] Performing multi-scale enhancement on the pre-processed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence;

[0151] The multi-scale enhanced sequence is input into the multi-layer graph convolution network of the wind speed prediction model, the values ​​of meteorological variables are used to represent nodes, the adjacency relationship between nodes is determined by the feature similarity between nodes, and the variable interaction features in the multi-scale enhanced sequence are extracted by frequency domain graph convolution in the hypervariable graph; the variable interaction features are used as node features of the spatial graph, meteorological stations are used to represent nodes, the adjacency relationship between nodes is determined by the geographical distance correlation between nodes, and the spatiotemporal features in the multi-scale enhanced sequence are extracted by residual graph convolution in the spatial graph;

[0152] The spatio-temporal features are subjected to inverted text embedding and low-rank sparse attention operation by a prediction module of the wind speed prediction model to obtain an attention matrix, and the attention matrix is input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features;

[0153] The variable fusion high-dimensional features are converted into wind speed prediction results by a fully connected layer of the wind speed prediction model.

[0154] Embodiment 4

[0155] The embodiment provides a computer readable storage medium, and computer programs / instructions are stored on the computer readable storage medium. When the computer programs / instructions are executed by a processor, the steps of the wind speed prediction method provided in the embodiment 1 are implemented.

[0156] After obtaining the multivariate weather time series collected by a plurality of weather stations in a target area, the multivariate weather time series are preprocessed, and the multivariate weather time series include values of a plurality of weather elements arranged in time sequence;

[0157] The preprocessed multivariate weather time series are input into the trained wind speed prediction model;

[0158] The preprocessed multivariate weather time series are subjected to multi-scale enhancement by a multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence;

[0159] The multi-scale enhanced sequence is input into a multi-layer graph convolution network of the wind speed prediction model, the values of the weather variables are used as nodes, the adjacency relationship between the nodes is determined according to the feature similarity between the nodes, the variable interaction features in the multi-scale enhanced sequence are extracted in a hyper-variable graph by frequency domain graph convolution, the variable interaction features are used as node features of a spatial graph, the weather stations are used as nodes, the adjacency relationship between the nodes is determined according to the geographical distance correlation between the nodes, and the spatio-temporal features in the multi-scale enhanced sequence are extracted in the spatial graph by residual graph convolution;

[0160] The spatio-temporal features are subjected to inverted text embedding and low-rank sparse attention operation by a prediction module of the wind speed prediction model to obtain an attention matrix, and the attention matrix is input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features;

[0161] The variable fusion high-dimensional features are converted into wind speed prediction results by a fully connected layer of the wind speed prediction model.

[0162] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0163] The present application is described in reference to the flowchart and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing system or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0164] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0165] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0166] The above description is only preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the technical principles of the present application, can make a number of improvements and variations, these improvements and variations should be considered as the protection scope of the present application.

Claims

1. A wind speed prediction method, characterized in that: include: Obtaining a multivariate meteorological time series collected by multiple meteorological stations in a target area and then preprocessing the series, wherein the multivariate meteorological time series includes values ​​of multiple meteorological elements arranged in time sequence; The preprocessed multivariate meteorological time series is input into the trained wind speed prediction model; Performing multi-scale enhancement on the pre-processed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence; The multi-scale enhanced sequence is input into the multi-layer graph convolutional network of the wind speed prediction model, the values ​​of meteorological variables are used to represent nodes, the adjacency relationship between nodes is determined by the feature similarity between nodes, and the variable interaction features in the multi-scale enhanced sequence are extracted through frequency domain graph convolution in the hypervariable graph; The variable interaction features are used as node features of the spatial graph, with weather stations representing nodes and the adjacency relationship between nodes determined by the correlation of geographical distances between nodes. In the spatial graph, the spatiotemporal features of the multi-scale enhanced sequence are extracted through residual graph convolution. Performing inverted text embedding and low-rank sparse attention operations on spatiotemporal features through the prediction module of the wind speed prediction model to obtain an attention matrix, and inputting the attention matrix into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features; The variable fusion high-dimensional features are converted into wind speed prediction results through the fully connected layer of the wind speed prediction model.

2. The wind speed prediction method according to claim 1, wherein: The meteorological elements include temperature, air pressure, wind speed, wind direction, dew point and rainfall.

3. The wind speed prediction method according to claim 1, wherein: The multi-scale enhanced sequence is obtained by performing multi-scale enhancement on the pre-processed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model, including: The preprocessed multivariate meteorological time series is layered downsampled and spliced ​​to obtain the downsampled splicing matrix; The downsampled splicing matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales; The spectral features at different time scales are fused to obtain a multi-scale enhanced sequence.

4. The wind speed prediction method according to claim 3, wherein: The layered downsampling is accomplished by layered average pooling; The downsampled concatenated matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales, including performing the following steps to extract spectral features at a single time scale at all time scales: Convert the downsampled stitching matrix at a single time scale to the frequency domain; The real and imaginary parts of the downsampled splicing matrix converted to the frequency domain are processed by the frequency domain multilayer perceptron respectively, and then restored to the time domain to obtain the spectral features at a single time scale.

5. The wind speed prediction method according to claim 1, wherein: The variable interaction features in the multi-scale enhanced sequence are extracted by frequency domain graph convolution in the hypervariable graph, which is performed by the following formula: ； in, represents a multi-scale enhanced sequence, represents the discrete Fourier transform, represents the inverse Fourier transform, Represents the transformation matrix in the set frequency domain, Represents variable interaction characteristics.

6. The wind speed prediction method according to claim 1, characterized in that: In the hypervariable graph, the adjacency matrix between any two nodes is expressed as: ； in, represents the adjacency matrix between two nodes in the hypervariable graph, represents the natural exponential function, Indicates the hypervariable graph The value of the meteorological variable represented by each node, Indicates the hypervariable graph The value of the meteorological variable represented by each node, Indicates the set scale parameter, Indicates the connection threshold between the set meteorological variables, Indicates finding the norm.

7. The wind speed prediction method according to claim 1, wherein: The spatiotemporal features in the multi-scale enhanced sequence are extracted by residual graph convolution in the spatial graph, which is performed by the following formula: ； in, represents the variable interaction characteristics; Represents the adjacency matrix between two nodes in the spatial graph, which is constructed based on the geographical locations of the weather stations corresponding to the two nodes; Represents the residual graph convolutional network; Represents the spatiotemporal features that include local variable dependencies and global spatial topology information in a multi-layer nested graph structure.

8. The wind speed prediction method according to claim 1, wherein: The attention matrix is ​​obtained by performing inverted text embedding and low-rank sparse attention operations on the spatiotemporal features, including: The complete spatiotemporal sequence of a meteorological variable in the spatiotemporal features at all time steps and all city weather stations is taken as a word unit. All words are combined into a matrix in the order of features and then transposed to obtain an inverted embedding sequence. The inverted embedding sequence is transformed into query matrix and key matrix through different linear transformations; Multiply the query matrix and key matrix with the corresponding set low-rank projection matrix to obtain the query matrix and key matrix after low-rank projection; According to the query matrix and key matrix after low-rank projection, the attention matrix is ​​calculated by the following formula: ； in, represents the attention matrix, represents the normalized exponential function, represents the query matrix after low-rank projection, represents the key matrix after low-rank projection, Represents a scaling factor used to prevent excessive or insufficient scaling when calculating inner products. Indicates transpose.

9. A wind speed prediction device, characterized in that: include: The multivariate acquisition module is configured to: acquire and preprocess a multivariate meteorological time series collected by multiple meteorological stations in a target area, wherein the multivariate meteorological time series includes values ​​of multiple meteorological elements arranged in time sequence; The wind speed prediction module is configured to: input the preprocessed multivariate meteorological time series into the trained wind speed prediction model; Performing multi-scale enhancement on the pre-processed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model to obtain a multi-scale enhanced sequence; The multi-scale enhanced sequence is input into the multi-layer graph convolutional network of the wind speed prediction model, the values ​​of meteorological variables are used to represent nodes, the adjacency relationship between nodes is determined by the feature similarity between nodes, and the variable interaction features in the multi-scale enhanced sequence are extracted through frequency domain graph convolution in the hypervariable graph; The variable interaction features are used as node features of the spatial graph, with weather stations representing nodes and the adjacency relationship between nodes determined by the correlation of geographical distances between nodes. In the spatial graph, the spatiotemporal features of the multi-scale enhanced sequence are extracted through residual graph convolution. Performing inverted text embedding and low sparse attention operations on spatiotemporal features through the prediction module of the wind speed prediction model to obtain an attention matrix, and inputting the attention matrix into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features; The variable fusion high-dimensional features are converted into wind speed prediction results through the fully connected layer of the wind speed prediction model.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the wind speed prediction method according to any one of claims 1 to 8 are implemented.

Images