A wind speed prediction method, device and storage medium

Abstract

The application 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 following steps: obtaining a multivariate weather time sequence collected by a plurality of weather stations in a target area and performing pretreatment, wherein the multivariate weather time sequence comprises values of a plurality of weather elements arranged in time sequence; inputting the pretreated multivariate weather time sequence into a trained wind speed prediction model to obtain a wind speed prediction result of a future set time step; wherein the wind speed prediction model introduces a multiscale decomposable hybrid module based on a frequency domain multilayer perception machine, a multilayer graph convolution network based on a hyper-variable graph and a spatial graph, and an Informer prediction module based on inverted text embedding. The application realizes the extraction of multivariate attributes, complex interaction relationships and multi-time scale characteristics by introducing the above modules, and realizes model lightweight, so as to solve the problems of low accuracy and low efficiency in the prior art.

Description

Technical Field

[0001] This invention relates to a wind speed prediction method, device, and storage medium, belonging to the field of wind speed prediction technology. Background Technology

[0002] Wind speed forecasting is a crucial component in modern weather forecasting, renewable energy dispatching (such as wind power), and aviation and shipping safety assurance, possessing significant research value and engineering implications. Traditional wind speed forecasting methods mainly include numerical weather prediction models, statistical regression models, and machine learning models based on single-station time series data. However, these methods often suffer from limitations in modeling capabilities, insufficient spatial relationship characterization, and sensitivity to non-stationary and high-noise data, making it difficult to meet the demands for high-precision wind speed forecasting.

[0003] In recent years, with the rapid development of deep learning technology, time series forecasting methods based on neural networks have gradually become a research hotspot. Among them, the Transformer (a neural network architecture based on an attention mechanism) has been widely used in meteorological element forecasting tasks due to its powerful global modeling capabilities. However, the standard Transformer model suffers from high computational overhead and insufficient ability to model spatial structures when processing long-term series. Furthermore, traditional spatial modeling methods often construct fixed topologies based on meteorological stations, neglecting the multivariate attributes and complex interactions of meteorological data, resulting in a weak ability of the forecasting model to learn the coupling relationships between variables.

[0004] Furthermore, wind speed data typically exhibits significant multi-timescale variations and high-frequency noise interference. Most existing models use fixed-scale inputs, making it difficult to fully extract variation patterns across different timescales, thus impacting prediction accuracy and the model's generalization ability.

[0005] In summary, existing technologies suffer from low accuracy and low efficiency in wind speed prediction. Summary of the Invention

[0006] The purpose of this invention is to provide a wind speed prediction method, device, and storage medium that can extract multivariate attributes and their complex interactions and multi-timescale features, and achieve lightweight modeling, so as to solve the problems of low accuracy and low efficiency in the wind speed prediction process of existing technologies.

[0007] To achieve the above objectives, the present invention employs the following technical solution:

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

[0009] After acquiring multivariate meteorological time series collected from multiple meteorological stations in the target area, preprocessing is performed. The multivariate meteorological time series includes the values ​​of multiple meteorological elements arranged in chronological order.

[0010] The preprocessed multivariate meteorological time series is input into the trained wind speed prediction model;

[0011] The multi-scale enhanced sequence is obtained by performing multi-scale enhancement on the preprocessed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model.

[0012] The multi-scale enhancement sequence is input into the multi-layer graph convolutional network of the wind speed prediction model. The values ​​of meteorological variables represent nodes, and the adjacency relationship between nodes is determined by the feature similarity between nodes. In the hypervariable graph, the variable interaction features in the multi-scale enhancement sequence are extracted by frequency domain graph convolution. The variable interaction features are used as node features in the spatial graph. Meteorological stations represent nodes, and the adjacency relationship between nodes is determined by the geographical distance correlation between nodes. In the spatial graph, the spatiotemporal features in the multi-scale enhancement sequence are extracted by residual graph convolution.

[0013] The wind speed prediction model's prediction module performs inverted text embedding and low-sparse attention operations on spatiotemporal features to obtain an attention matrix. The attention matrix is ​​then input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features.

[0014] The fully connected layer of the wind speed prediction model fuses high-dimensional features into wind speed prediction results.

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

[0016] Furthermore, the multi-scale enhancement of the preprocessed multivariate meteorological time series is performed by the multi-scale decomposable hybrid module of the wind speed prediction model to obtain the multi-scale enhanced sequence, which includes: performing hierarchical downsampling and splicing on the preprocessed multivariate meteorological time series to obtain the downsampling splicing matrix.

[0017] The downsampled splicing matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales;

[0018] Multi-scale enhanced sequences are obtained by fusing spectral features at different time scales.

[0019] Furthermore, the hierarchical downsampling is accomplished through hierarchical average pooling;

[0020] The downsampled concatenation matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales, including the following steps for extracting spectral features at a single time scale across all time scales:

[0021] Transform the downsampled concatenation matrix at a single time scale to the frequency domain;

[0022] The real and imaginary parts of the downsampled splicing matrix transformed to the frequency domain are processed by a frequency domain multilayer perceptron and then restored to the time domain to obtain the spectral features at a single time scale.

[0023] Furthermore, variable interaction features in the multi-scale enhanced sequence are extracted through frequency domain graph convolution in the hypervariable graph, using the following formula:

[0024] ;

[0025] in, Indicates multi-scale enhancement sequences, Represents the Discrete Fourier Transform. Indicates the inverse Fourier transform. This represents the transformation matrix in the defined frequency domain. This indicates the interaction characteristics of variables.

[0026] Furthermore, in the hypervariable graph, the expression for the adjacency matrix between any two nodes is:

[0027] ;

[0028] in, This represents the adjacency matrix between two nodes in the hypervariable graph. This represents the natural exponential function. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. This indicates the set scale parameters. This represents the threshold value set between meteorological variables. This indicates a search for the norm.

[0029] Furthermore, spatiotemporal features in the multi-scale enhanced sequence are extracted from the spatial map through residual map convolution, using the following formula:

[0030] ;

[0031] in, Indicates the interaction characteristics of variables; The adjacency matrix representing the two nodes in the spatial graph is constructed based on the geographical locations of the weather stations corresponding to the two nodes. Represents a residual graph convolutional network; This represents the spatiotemporal characteristics that include local variable dependencies and global spatial topological information in a multi-layered nested graph structure.

[0032] Furthermore, the attention matrix is ​​obtained by performing inverted text embedding and low-rank sparse attention operations on the spatiotemporal features:

[0033] The complete spatiotemporal sequence of a meteorological variable in all time steps and all cities is taken as a word. All words are combined into a matrix according to the feature order and then transposed to obtain the inverted embedding sequence.

[0034] The query matrix and key matrix are obtained by performing different linear transformations on the inverted embedded sequence;

[0035] Multiply the query matrix and key matrix with the corresponding low-rank projection matrix to obtain the low-rank projected query matrix and key matrix;

[0036] Based on the query matrix and key matrix after low-rank projection

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

[0038] ;

[0039] in, Represents the attention matrix. Represents the normalized exponential function, This represents the query matrix after low-rank projection. This represents the key matrix after low-rank projection. This indicates a scaling factor used to prevent excessively large or small values ​​when calculating the inner product. This indicates transpose.

[0040] Secondly, the present invention provides a wind speed prediction device, comprising:

[0041] The multivariate acquisition module is configured to: acquire multivariate meteorological time series collected by multiple meteorological stations in the target area and then perform preprocessing, wherein the multivariate meteorological time series includes the values ​​of multiple meteorological elements arranged in chronological order;

[0042] The wind speed prediction module is configured to input the preprocessed multivariate meteorological time series into the trained wind speed prediction model;

[0043] The multi-scale enhanced sequence is obtained by performing multi-scale enhancement on the preprocessed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model.

[0044] The multi-scale enhancement sequence is input into the multi-layer graph convolutional network of the wind speed prediction model. The values ​​of meteorological variables represent nodes, and the adjacency relationship between nodes is determined by the feature similarity between nodes. In the hypervariable graph, the variable interaction features in the multi-scale enhancement sequence are extracted by frequency domain graph convolution. The variable interaction features are used as node features in the spatial graph. Meteorological stations represent nodes, and the adjacency relationship between nodes is determined by the geographical distance correlation between nodes. In the spatial graph, the spatiotemporal features in the multi-scale enhancement sequence are extracted by residual graph convolution.

[0045] The wind speed prediction model's prediction module performs inverted text embedding and low-rank sparse attention operations on the spatiotemporal features to obtain an attention matrix. The attention matrix is ​​then input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features.

[0046] The fully connected layer of the wind speed prediction model fuses high-dimensional features into wind speed prediction results.

[0047] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the wind speed prediction method described in any one of the first aspects.

[0048] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0049] This invention provides a wind speed prediction method, device, and storage medium. By introducing a multi-scale decomposable hybrid module based on a frequency-domain multilayer perceptron, it significantly enhances the model's ability to model different frequency components in non-stationary wind speed time series and suppresses the influence of high-frequency noise on the prediction results. Simultaneously, by constructing a hypervariable graph and combining it with a graph convolutional network, dual modeling at both the variable level and spatial level is achieved, enabling efficient extraction of complex dependencies in multi-site, multivariate wind speed data. In the prediction stage, an improved Informer structure is introduced, employing an inverted embedding mechanism and a low-rank sparse attention mechanism to significantly improve the ability to capture long-term dependency information and reduce model computational complexity. Compared with existing wind speed prediction methods, this invention demonstrates significant advantages in accuracy, stability, and computational efficiency, particularly exhibiting good generalization ability and practicality in multivariate interactive modeling and spatiotemporal fusion prediction tasks.

[0050] This invention is applicable to applications such as ultra-short-term wind speed forecasting for wind farms, dynamic dispatching of power systems, and meteorological early warning, and has significant engineering promotion value. Attached Figure Description

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

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

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

[0054] Figure 4 This is a schematic diagram of the processing flow of the multi-scale decomposable hybrid module provided by the present invention for its input. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, and should not be used to limit the scope of protection of the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment provides a wind speed prediction method, including:

[0058] After acquiring multivariate meteorological time series collected from multiple meteorological stations in the target area, preprocessing is performed. The multivariate meteorological time series includes the values ​​of multiple meteorological elements arranged in chronological order.

[0059] The preprocessed multivariate meteorological time series is input into the trained wind speed prediction model;

[0060] Multiscale enhancement of the preprocessed multivariate meteorological time series was performed by the multiscale decomposable hybrid module of the wind speed prediction model to obtain the multiscale enhanced sequence.

[0061] The multi-scale augmentation sequence is input into a multi-layer graph convolutional network of the wind speed prediction model. The values ​​of meteorological variables represent nodes, and the adjacency relationship between nodes is determined by the feature similarity between nodes. In the hypervariable graph, the variable interaction features in the multi-scale augmentation sequence are extracted by frequency domain graph convolution. The variable interaction features are used as node features in the spatial graph. The meteorological station represents nodes, and the adjacency relationship between nodes is determined by the geographical distance correlation between nodes. In the spatial graph, the spatiotemporal features in the multi-scale augmentation sequence are extracted by residual graph convolution.

[0062] The wind speed prediction model's prediction module performs inverted text embedding and low-rank sparse attention operations on spatiotemporal features to obtain an attention matrix. The attention matrix is ​​then input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features.

[0063] The fully connected layer of the wind speed prediction model fuses high-dimensional features to transform the variables into wind speed prediction results.

[0064] This invention significantly enhances the model's ability to model different frequency components in non-stationary wind speed time series by introducing a multi-scale decomposable hybrid module based on a frequency-domain multilayer perceptron, thus suppressing the impact of high-frequency noise on prediction results. Simultaneously, it constructs a hypervariable graph and combines it with a graph convolutional network to achieve dual modeling at both the variable and spatial levels, enabling efficient extraction of complex dependencies in multi-site, multivariate wind speed data. In the prediction stage, an improved Informer structure is introduced, employing an inverted embedding mechanism and a low-rank sparse attention mechanism to significantly improve the ability to capture long-term dependency information and reduce model computational complexity. Compared with existing wind speed prediction methods, this invention demonstrates significant advantages in accuracy, stability, and computational efficiency, particularly exhibiting good generalization ability and practicality in multivariate interactive modeling and spatiotemporal fusion prediction tasks.

[0065] Example 2

[0066] according to Figure 2 The technical concept shown is based on Figure 3 The flowchart shown illustrates a wind speed prediction method, which includes the following steps one through six.

[0067] The first step is to normalize the dataset.

[0068] The datasets used in this invention are from two actual meteorological datasets from Region 1 and Region 2. The actual meteorological dataset from Region 2 contains four meteorological elements (temperature, air pressure, wind speed, and wind direction) from five cities, with data recorded hourly. The actual meteorological dataset from Region 1 contains meteorological data from seven cities, including six meteorological elements: wind speed, wind direction, temperature, dew point, air pressure, and rainfall. Both actual meteorological datasets have a [format missing]. ,in, Indicates the quantity of meteorological elements. The number of time steps. The number of cities is represented by [value]. The data from the two actual meteorological datasets are arranged chronologically into multivariate meteorological time series.

[0069] To perform data preprocessing, min-max normalization was applied to each dataset, scaling all data to the range [0,1]. The normalization formula is as follows:

[0070] (1);

[0071] in, This represents the raw data in the actual meteorological dataset. and These are the minimum and maximum values ​​in the original data, respectively. This 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, to effectively model the multi-timescale features in wind speed data, this invention designs a multi-scale decomposable hybrid module based on a frequency domain multilayer perceptron, comprising the following three parts:

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

[0075] The input multivariate meteorological time series is downsampled using hierarchical average pooling to obtain a multi-timescale representation. Let the input of the multi-scale decomposable hybrid module be... , Represents a shape as Given a dataset, let the maximum number of downsampling layers be... Then, multiple time scales are represented as:

[0076] (2);

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

[0078] For multiple sequences obtained through downsampling, a downsampling concatenation matrix is ​​obtained by concatenating them along the channel dimension. The downsampled concatenation matrix is ​​then used as the input for the next step.

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

[0080] The downsampled concatenation matrix at each time scale is transformed to the frequency domain using a Discrete Fourier Transform (DFT):

[0081] (3);

[0082] in, This represents the downsampling concatenation matrix transformed to the frequency domain. Represents the Discrete Fourier Transform. It is a downsampling concatenation matrix.

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

[0084] (4);

[0085] (5);

[0086] in, This represents the real part of the downsampled concatenation matrix transformed to the frequency domain. This represents the imaginary part of the downsampling concatenation matrix after transformation to the frequency domain. and This is the weight matrix. and For bias terms, For activation function, This represents the result of processing the real part of the downsampled concatenation matrix transformed to the frequency domain using a multilayer perceptron. This represents the result of processing the imaginary part of the downsampled concatenation matrix transformed to the frequency domain using a multilayer perceptron.

[0087] Depend on and Composed of a downsampled splicing matrix processed by a multilayer perceptron Then, the inverse Fourier transform is used to transform the downsampled concatenation matrix processed by the multilayer perceptron. Restore to the time domain:

[0088] (6);

[0089] in, This represents the downsampled concatenation matrix processed by the multilayer perceptron. Indicates the inverse Fourier transform. It represents the spectral characteristics at a single time scale in the time domain.

[0090] (3) Residual fusion

[0091] To better integrate spectral features from different time scales and avoid feature loss, this invention employs the following residual fusion mechanism:

[0092] (7);

[0093] in, This represents the downsampling concatenation matrix. Represents the spectral characteristics at a single time scale in the time domain. Represents the Discrete Fourier Transform. This represents a multi-scale augmentation sequence, which is the input to a multi-layer graph convolutional network.

[0094] The third step is to extract spatiotemporal features from the multi-layer graph convolutional network in the wind speed prediction model.

[0095] In wind speed forecasting tasks, to effectively capture the complex spatial dependencies between different weather stations and between multiple variables within the same weather station, this invention proposes a multi-layer graph convolutional network, which includes two layers: hypervariable graph modeling and spatial graph convolutional modeling between stations. It achieves efficient spatiotemporal modeling capabilities through nested fusion.

[0096] Traditional modeling methods using weather stations as graph nodes can only represent the spatial relationships between stations, making it difficult to depict the interactions between different meteorological variables within the same station. Therefore, this invention introduces a Hypervariate Graph Network (HGN), where multiple meteorological variables (such as wind speed, temperature, and humidity) are treated as independent graph nodes within each weather station, establishing a graph structure between variables to enhance the expressive power of the modeling.

[0097] In a hypervariable graph, the adjacency relationship between any two nodes is controlled by their feature similarity, and the adjacency matrix... The definition is as follows:

[0098] (8);

[0099] in, This represents the adjacency matrix between two nodes in the hypervariable graph. This represents the natural exponential function. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. This indicates the set scale parameters. This represents the threshold value set between meteorological variables. This indicates a search for the norm.

[0100] After constructing the hypervariable graph, the frequency domain graph convolution method is used to extract variable interaction features in the multi-scale enhanced sequence (the output features of variable dependencies are fused after frequency domain graph convolution) to reduce computational complexity and improve modeling efficiency.

[0101] (9);

[0102] in, Indicates multi-scale enhancement sequences, Represents the Discrete Fourier Transform. Indicates the inverse Fourier transform. This represents the transformation matrix in the defined frequency domain. This indicates the interaction characteristics of variables.

[0103] After completing the hypervariable graph modeling within the weather station, this invention further constructs a spatial graph structure between weather stations and uses Residual Graph Convolutional Network (RGCN) for modeling to achieve information fusion from local (within the weather station) to global (between weather stations).

[0104] First, output the aforementioned hypervariable graph. As a node characteristic, the nodes of the spatial graph are defined as weather stations, and the connection relationships between nodes are determined by the adjacency matrix. Representation (constructed based on geographic distance correlation). The residual map convolution operation is performed on the spatial map as follows:

[0105] A spatial graph structure is established using weather stations as nodes, and residual graph convolution is performed:

[0106] (10);

[0107] in, This represents the output of the hypervariable graphical network. This represents the adjacency matrix between two nodes in a spatial graph. The adjacency matrix between two nodes is constructed based on the geographical locations of the weather stations corresponding to the two nodes. This represents a residual graph convolutional network. This represents the spatiotemporal features that include local variable dependencies and global spatial topological information within a multi-layered nested graph structure. As input to the prediction module.

[0108] The fourth step involves using an Informer prediction module based on inverted embedding (text embedding) to predict wind speed at a preset future time step. Informer is a variant of Transformer.

[0109] After completing multi-scale temporal feature extraction and multi-layer graph convolutional modeling, this invention further designs an efficient time series prediction mechanism, namely, an Informer prediction module based on inverted embedding, named iInformer, which is used to generate wind speed prediction results for future time steps. This module combines the inverted text embedding idea of ​​iTransformer with the sparse attention mechanism of Informer, effectively modeling long-term dependencies between multiple variables while reducing computational complexity, thus improving prediction accuracy and efficiency.

[0110] The data organization method of inverted text embedding is as follows: Traditional Transformer models treat each time step as an input lexical unit. When processing multivariate data, variables with different physical meanings need to be projected into the same embedding space, which may lead to confusion between variables; at the same time, this structure has high computational complexity when processing long-term series. This invention adopts an inverted text embedding mechanism, treating the time series of each variable as an independent lexical unit, thereby allowing the attention mechanism to directly act between variables. The original input data is... ,in This represents the characteristic dimension of each city; Indicates the number of time steps; This indicates the number of cities (weather stations).

[0111] Traditional Transformers use a time-based approach for text embedding, which is achieved through the following formula:

[0112] (11);

[0113] in, This represents the embedding sequence obtained by text embedding using a traditional Transformer. Represents all features at the first time step. Represents all features at the second time step. It represents all features at the Nth time step. In the traditional Transformer, a time step is used as a lexicon, which includes N variables.

[0114] In the iInformer provided in this embodiment, variable-based inverted text embedding is used, which is achieved through the following formula:

[0115] (12);

[0116] in, Indicates an inverted embedding sequence. This represents the complete sequence of the first feature across all time steps. This represents the complete sequence of the second feature across all time steps. Indicates the first In iInformer, the complete time series of a meteorological variable in the spatiotemporal features across all time steps is treated as a lexical unit, thereby enabling the attention mechanism to capture long-term dependencies between variables and reduce redundant computations at time steps.

[0117] Building upon the inverted embedding described above, the Transformer structure is further optimized by introducing a kernel-based low-rank sparse attention mechanism to reduce computational burden. The attention matrix is ​​calculated using this kernel-based low-rank sparse attention mechanism:

[0118] The query matrix is ​​obtained by performing different linear transformations on the inverted embedded sequences. Bond matrix ;

[0119] Perform a low-rank projection on the query matrix and the key matrix:

[0120] (13);

[0121] (14);

[0122] in, Represents the query matrix. Represents the key matrix, and This represents the defined low-rank projection matrix. This represents the query matrix after low-rank projection. This represents the key matrix after low-rank projection;

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

[0124] (15);

[0125] in, Represents the attention matrix. Represents the normalized exponential function, This indicates a scaling factor used to prevent excessively large or small values ​​when calculating the inner product. This indicates transpose.

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

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

[0128] To obtain the target predicted value, this invention introduces a fully connected layer at the end of the model, mapping the internal representation to the wind speed prediction space to generate the final wind speed prediction result for a given future time step. This fully connected layer can be formally represented as:

[0129] (16);

[0130] in, This indicates the wind speed forecast result. Represents the ReLU activation function. and These are the parameters for the fully connected network.

[0131] Step 5: Model validation and training.

[0132] After the wind speed prediction model is built, it needs to be trained and validated to ensure its effectiveness and generalization ability in wind speed prediction tasks.

[0133] To verify the effectiveness of the proposed model in wind speed prediction, this invention uses two real datasets, Region 2 and Region 1, and employs Mean Absolute Error (MAE) and Mean Squared Error (MSE) as evaluation metrics. The definitions of MAE and MSE are as follows:

[0134] (17);

[0135] (18);

[0136] in, Represents the first in the real dataset The true value of each sample Represents the first in the real dataset The predicted value for each sample, This represents the total number of samples in the real dataset.

[0137] Regarding the loss function, this invention selects MSE as the loss function and Adam (Adaptive Moment Estimation, an optimization algorithm based on gradient descent) as the optimizer, with an initial learning rate set to 0.0001. The batch size is set to 64, the number of training epochs is 50, and an early stopping strategy is adopted for the validation set.

[0138] Step 6: Wind speed prediction test.

[0139] To further verify the practicality and superiority of the wind speed prediction model (named MS-iInformer-GCNs) proposed in this invention for wind speed prediction tasks, two actual meteorological datasets were selected, from multiple meteorological stations in Region 2 and Region 1, respectively. In the experiments, the method provided by this invention was compared with methods using mainstream sequence models for wind speed prediction, including recurrent neural network structures (GRU), Transformer, and its improved version Informer. Wind speed prediction tasks were performed at different time spans, and the prediction accuracy and stability of each model were compared.

[0140] On a real meteorological dataset from Region 2, the test task was set to predict wind speed for the next 6, 12, 18, and 24 hours. The experimental results shown in Table 1 indicate that MS-iInformer-GCNs had the lowest MAE and MSE for all prediction durations. Specifically, the MAEs for predicting the next 6, 12, 18, and 24 hours were 1.155, 1.408, 1.523, and 1.624, respectively, while the MSEs were 2.370, 3.383, 3.899, and 4.371, respectively. In contrast, the traditional GRU had an MAE higher than 1.6 for all time periods, and an MSE exceeding 5.9 for 24 hours, significantly higher than the model provided in this invention. Although Transformer and Informer show significant improvements over GRU, they still tend to increase errors in long-term predictions, especially in 24-hour predictions, where their MSEs reach 4.978 and 4.950 respectively, while the model of this invention achieves 4.371, indicating that it has better stability and accuracy in long-term series modeling.

[0141] Table 1 - MAE and MSE tables for wind speed prediction tasks from actual meteorological datasets of Region 2.

[0142]

[0143] On a regional city dataset, prediction tasks were set for multiple time spans from 2 to 10 hours, and model performance was evaluated from both MAE and MSE perspectives. The experimental results shown in Table 1 demonstrate that MS-iInformer-GCNs continues to exhibit superior prediction performance. The MAEs for 2h, 4h, 6h, 8h, and 10h are 7.38, 8.41, 9.35, 10.06, and 10.64, respectively, while the MSEs are 96.59, 125.10, 155.03, 180.41, and 201.76, respectively, comprehensively outperforming other models. Especially in the long-term prediction of 10 hours, GRU achieves an MSE as high as 382.54, and Informer and Transformer also exceed 220, while the model provided by this invention achieves an MSE of 201.76, demonstrating a strong generalization ability for modeling complex spatiotemporal dependencies.

[0144] Table 2 - MAE and MSE tables for wind speed prediction tasks from actual meteorological datasets of 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 multivariate meteorological time series collected by multiple meteorological stations in the target area and then perform preprocessing, wherein the multivariate meteorological time series includes the values ​​of multiple meteorological elements arranged in chronological order;

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

[0150] The multi-scale enhanced sequence is obtained by performing multi-scale enhancement on the preprocessed multivariate meteorological time series through the multi-scale decomposable hybrid module of the wind speed prediction model.

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

[0152] The wind speed prediction model's prediction module performs inverted text embedding and low-rank sparse attention operations on the spatiotemporal features to obtain an attention matrix. The attention matrix is ​​then input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features.

[0153] The fully connected layer of the wind speed prediction model fuses high-dimensional features into wind speed prediction results.

[0154] Example 4

[0155] This embodiment provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the wind speed prediction method provided in Embodiment 1:

[0156] After acquiring multivariate meteorological time series collected from multiple meteorological stations in the target area, preprocessing is performed. The multivariate meteorological time series includes the values ​​of multiple meteorological elements arranged in chronological order.

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

[0158] Multiscale enhancement of the preprocessed multivariate meteorological time series was performed by the multiscale decomposable hybrid module of the wind speed prediction model to obtain the multiscale enhanced sequence.

[0159] The multi-scale augmentation sequence is input into a multi-layer graph convolutional network of the wind speed prediction model. The values ​​of meteorological variables represent nodes, and the adjacency relationship between nodes is determined by the feature similarity between nodes. In the hypervariable graph, the variable interaction features in the multi-scale augmentation sequence are extracted by frequency domain graph convolution. The variable interaction features are used as node features in the spatial graph. The meteorological station represents nodes, and the adjacency relationship between nodes is determined by the geographical distance correlation between nodes. In the spatial graph, the spatiotemporal features in the multi-scale augmentation sequence are extracted by residual graph convolution.

[0160] The wind speed prediction model's prediction module performs inverted text embedding and low-rank sparse attention operations on spatiotemporal features to obtain an attention matrix. The attention matrix is ​​then input into a layer normalization and feedforward neural network to obtain variable fusion high-dimensional features.

[0161] The fully connected layer of the wind speed prediction model fuses high-dimensional features to transform the variables into wind speed prediction results.

[0162] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0163] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations 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 processor, 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, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0166] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

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

2. The wind speed prediction method according to claim 1, characterized in that, 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, characterized in that, The preprocessed multivariate meteorological time series is enhanced at multiple scales using the multiscale decomposable hybrid module of the wind speed prediction model to obtain a multiscale enhanced sequence, including: The preprocessed multivariate meteorological time series is subjected to stratified downsampling and splicing to obtain a downsampling splicing matrix; The downsampled splicing matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales; Multi-scale enhanced sequences are obtained by fusing spectral features at different time scales.

4. The wind speed prediction method according to claim 3, characterized in that, The hierarchical downsampling is accomplished through hierarchical average pooling; The downsampled concatenation matrix is ​​processed by a frequency domain multilayer perceptron to obtain spectral features at different time scales, including the following steps for extracting spectral features at a single time scale across all time scales: Transform the downsampled concatenation matrix at a single time scale to the frequency domain; The real and imaginary parts of the downsampled splicing matrix transformed to the frequency domain are processed by a frequency domain multilayer perceptron 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, characterized in that, In the hypervariable graph, variable interaction features in the multi-scale enhanced sequence are extracted through frequency domain graph convolution, using the following formula: ； in, Indicates multi-scale enhancement sequences, Represents the Discrete Fourier Transform. Indicates the inverse Fourier transform. This represents the transformation matrix in the defined frequency domain. This indicates the interaction characteristics of variables.

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, This represents the adjacency matrix between two nodes in the hypervariable graph. This represents the natural exponential function. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. In the hypervariable graph, the first... The values ​​of the meteorological variables represented by each node. This indicates the set scale parameters. This represents the threshold value set between meteorological variables. This indicates a search for the norm.

7. The wind speed prediction method according to claim 1, characterized in that, Spatiotemporal features in multi-scale augmented sequences are extracted through residual map convolution in the spatial map, using the following formula: ； in, Indicates the interaction characteristics of variables; The adjacency matrix representing the two nodes in the spatial graph is constructed based on the geographical locations of the weather stations corresponding to the two nodes. Represents a residual graph convolutional network; This represents the spatiotemporal characteristics that include local variable dependencies and global spatial topological information in a multi-layered nested graph structure.

8. The wind speed prediction method according to claim 1, characterized in that, An attention matrix is ​​obtained by performing inverted text embedding and low-rank sparse attention operations on spatiotemporal features, including: The complete spatiotemporal sequence of a meteorological variable in all time steps and all urban meteorological stations is taken as a word. All words are combined into a matrix according to the feature order and then transposed to obtain the inverted embedding sequence. The query matrix and key matrix are obtained by performing different linear transformations on the inverted embedded sequence; Multiply the query matrix and key matrix with the corresponding low-rank projection matrix to obtain the low-rank projected query matrix and key matrix; The attention matrix is ​​calculated using the following formula, based on the query matrix and key matrix after low-rank projection: ； in, Represents the attention matrix. Represents the normalized exponential function, This represents the query matrix after low-rank projection. This represents the key matrix after low-rank projection. This indicates a scaling factor used to prevent excessively large or small values ​​when calculating the inner product. This indicates transpose.

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

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

Images