Surface wind speed prediction method based on multi-source data and spatial-temporal feature fusion

Abstract

The invention provides an earth surface wind speed prediction method based on multi-source data and spatio-temporal feature fusion, and belongs to the technical field of meteorological prediction and computer cross, and the method comprises the following steps: S1, carrying out the preprocessing of a data set; s2, utilizing a TC-ResNet module to extract spatial and temporal characteristics of the wind speed data; s3, a time sequence dynamic mode of the data is mined through a TE-Liquid Time-Constance Networks module, and a time sequence dynamic mode of the data is mined through a TE-Liquid Time-Constance Networks module; s4, splicing features through a gating fusion mechanism, calculating weights, and inputting the weights into MLP for surface wind speed prediction; s5, Bayesian optimization hyper-parameters are adopted, and in combination with an adaptive moment estimation (AdamW) optimizer with weight attenuation, cosine annealing learning rate scheduling and smooth L1 (SmoonL1Loss) loss function optimization model training are carried out; and S6, performing short-term surface wind speed prediction on the test set by using the trained model. The method has the advantage that the accuracy of surface wind speed prediction under the complex terrain is improved.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting and computer science, and in particular to a method for predicting surface wind speed based on the fusion of multi-source data and spatiotemporal features. Background Technology

[0002] With the global energy structure transformation and the development of meteorological forecasting technology, short-term wind speed forecasting is becoming increasingly important for wind energy resource assessment and stable power grid operation. Current numerical weather prediction (NWP) relies on global data and atmospheric dynamic equations for solving problems. While it is the mainstream method for forecasting meteorological elements, the chaotic nature and uncertainty of the atmospheric system make it difficult to accurately capture local micro-meteorological factors such as surface, land-sea thermal conditions, and topography, leading to significant biases in wind speed forecasts. Although traditional statistical methods and machine learning models attempt to improve forecast accuracy, most models struggle to balance the synergistic extraction of spatiotemporal characteristics and element correlations, and they are insufficient in representing near-surface atmospheric motion under the influence of complex terrain.

[0003] In recent years, deep learning methods have been widely used in meteorology due to their superior nonlinear fitting and feature fusion capabilities, especially hybrid spatiotemporal prediction models which have begun to show potential in wind speed prediction. However, while such models can handle both spatial and temporal features simultaneously, they struggle to capture complex dynamic relationships, do not fully utilize multi-source data, and have poor adaptability to complex terrain and extreme wind speed events.

[0004] Solving the aforementioned problems is the key to this invention. This invention deeply integrates temporal embedding (TE) technology with a dual-branch architecture, addressing the insufficient ability of traditional models to distinguish wind speed differences under similar terrain, and mitigating noise interference from multi-source data through a convolutional block attention mechanism (CBAM). This method, through a spatiotemporal feature fusion architecture and adaptive gating mechanism, overcomes the shortcomings of existing technologies in predicting surface wind speed in complex terrain, providing technical support for the efficient development of wind energy resources and the optimization of energy systems. Summary of the Invention

[0005] The purpose of this invention is to provide a surface wind speed prediction method based on the fusion of multi-source data and spatiotemporal features, which aims to improve the accuracy of surface wind speed prediction under complex terrain.

[0006] The core idea of ​​this invention is as follows: This invention proposes a surface wind speed prediction method based on the fusion of multi-source data and spatiotemporal features. By integrating meteorological parameters from the THORPEX Interactive Global Encyclopedia (TIGGE) project, topographic data from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3), and temporal parameters, a dual-branch architecture is constructed to extract spatiotemporal dynamic features. Combined with a Convolutional Block Attention (CBAM) mechanism, key influencing factors are focused to generate a spatiotemporally coupled prediction model. Temporal embedding (TE) technology enhances the model's perception of wind speed periodicity, and a gated fusion mechanism adaptively fuses the dual-branch features, improving the accuracy and reliability of surface wind speed prediction under complex terrain. This enables the method to effectively correct biases in numerical weather prediction, providing precise data support for wind energy resource assessment and wind farm scheduling, and reducing the operational risks of energy systems.

[0007] This invention is achieved through the following measures: a surface wind speed prediction method based on the fusion of multi-source data and spatiotemporal features, comprising the following steps:

[0008] 1.1: The meteorological parameters, topographic data of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3), and time parameters of the THORPEX Interactive Global Encyclopedia (TIGGE) project were preprocessed. Outliers were removed by spatiotemporal linear interpolation and 3σ rule and aligned to a unified grid. The gradient boosting decision tree (CatBoost) model was used to select a specific number of parameters from 27 meteorological parameters and 4 topographic parameters that were most correlated with the wind speed data in the Japan 55-Year Reanalysis (JRA-55). The training set, validation set, and test set were divided according to the year.

[0009] 1.2: Utilize the ResNet module, incorporate Temporal Embedding (TE) technology, and combine it with Convolutional Block Attention (CBAM) to extract the spatiotemporal features of wind speed data;

[0010] 1.3: By introducing the Liquid Time-Constant Networks module, the Temporal Embedding (TE) technique is introduced, and the time series dynamic patterns of wind speed data are mined using gating mechanisms and meta-learning.

[0011] 1.4: By splicing dual-branch features through a gating fusion mechanism, calculating the fusion weights, and then inputting them into the MLP, surface wind speed prediction is achieved;

[0012] 1.5: Bayesian optimization of hyperparameters is adopted, combined with the weighted decay adaptive moment estimation (AdamW) optimizer, cosine annealing learning rate and smooth L1 loss function to optimize model training;

[0013] 1.6: Use the trained surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features to perform short-term surface wind speed prediction on the test set and output the prediction results.

[0014] Step 1.1 includes the following steps:

[0015] 2.1: Missing values ​​were processed for meteorological parameters (including 10-meter wind speed components u10 and v10, topography orog, soil moisture sm, etc.) from the THORPEX Interactive Global Encyclopedia (TIGGE) project, topographic data (ASTER GDEM V3) (elevation, relief, slope, aspect), and time parameters (year, month, day, hour, season) from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (ASTER GDEM V3). Linear interpolation was used in the time dimension, and linear interpolation was first performed in the spatial dimension, followed by nearest neighbor interpolation.

[0016] 2.2: The 3σ rule was used to remove outliers that deviated from the mean by more than 3 times the standard deviation, and all data were uniformly interpolated to a 0.25°×0.25° grid to match the four time points of 06:00, 12:00, 18:00 and 24:00 each day;

[0017] 2.3: The gradient boosting decision tree (CatBoost) model was used to calculate feature importance. Through a two-stage feature selection method, eight key meteorological parameters were selected from the meteorological parameters of the THORPEX interactive global ensemble (TIGGE) project, and three key terrain parameters were selected from the terrain data of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3). Finally, the training set, validation set and test set were divided according to the year.

[0018] In step 1.2, a residual network (TC-ResNet) branch architecture incorporating temporal embedding (TE) and convolutional block attention (CBAM) is used to extract the spatiotemporal features of meteorological data from the dataset processed in step 1.1. This includes the following steps:

[0019] 3.1: Spatial features are extracted using the ResNet architecture, and a Convolutional Block Attention (CBAM) mechanism is introduced. Channel attention is calculated using global average pooling to determine channel weights, as shown in the following formula:

[0020]

[0021] α=σ(W1·ReLU(W0·AvgPool(x)))

[0022] Where, AvgPool(x) c This represents the result of global average pooling on the c-th channel of the input feature map x, where i represents the latitude index of the feature map (ranging from 1 to 48), j represents the longitude index of the feature map (ranging from 1 to 96), and x... c (i,j) represents the feature value located at grid point (i,j) in the c-th channel, α represents the generated channel attention weight matrix, σ represents the sigmoid activation function, W0 represents the dimension reduction convolution kernel, W1 represents the dimension increase convolution kernel, and ReLU represents the linear rectified activation function.

[0023] 3.2: Spatial attention extracts spatial cues through channel-dimensional average pooling and max pooling, as shown in the following formula:

[0024]

[0025] β=σ(Conv 7×7 (Concat(F avg ,F max )))

[0026] Among them, F avg (i,j) represents the result at grid point (i,j) after average pooling of the feature map along the channel dimension, 11 represents the total number of channels of the input feature, F max (i,j) represents the result at grid point (i,j) after max pooling the feature map along the channel dimension, β represents the generated spatial attention weight map, and Conv 7×7 This represents a 7×7 convolution operation that performs spatial feature extraction and dimensionality compression on the concatenated feature map. Concat(F) avg ,F max This indicates that the two results of channel pooling are concatenated along the channel dimension.

[0027] 3.3: A three-layer fully connected network is used to map the time parameters into a 56-dimensional time embedding vector, enhancing the ability to extract features related to time dependencies. The formula is as follows:

[0028] TE(t)=W4·ReLU(W3·ReLU(W2·t))

[0029] Where TE(t) represents the output of the time embedding, t represents the set of time-related parameters of the input, W2 represents the weight matrix of the first fully connected network, W3 represents the weight matrix of the second fully connected network, and W4 represents the weight matrix of the third fully connected network.

[0030] In step 1.3, the dynamic patterns of the time series of wind speed data in the dataset processed in step 1.1 are mined using a branch architecture of TE-LiquidTime-Constant Networks, which incorporates temporal embedding (TE) technology, and employs gating mechanisms and meta-learning. Specifically, this includes the following steps:

[0031] 4.1: Modeling the continuity of time series using ordinary differential equations based on Liquid Time-Constant Networks, and performing numerical integration using the Runge-Kutta RK4 method;

[0032] 4.2: Based on the gating mechanism that integrates time characteristics and input data, the formula is as follows:

[0033] g t =σ(W g [x t ;e′ t ]+b g ),f t =g t ⊙x t +(1-g t )⊙e′ t

[0034] Where g t W represents the gating weight at time step t. g The weight matrix represents the gating mechanism, used to perform a linear transformation on the concatenated features, x t The input data for time step t, e t ′ represents the temporal embedding feature after processing by a multilayer perceptron, b g The bias term representing the gating mechanism is a parameter learned during model training, used to adjust the baseline value of the linear transformation, f. t ⊙ represents the output of the gating mechanism;

[0035] 4.3: Time-varying weights and biases are generated from temporal features through a meta-learning mechanism to dynamically modulate the parameters of the ordinary differential equation (ODE) function;

[0036] 4.4: Adjusting the fourth-order (RK4) integral step size based on time characteristics enhances the ability to model non-stationary sequences. The formula is as follows:

[0037]

[0038] Wherein, Δt(e′) t ) represents the time-embedded feature e′ tThe calculated fourth-order (RK4) integral step size, where A and B represent hyperparameters used to adjust the overall scale and offset of the step size function, ensures that the step size is a reasonably positive value. b represents the transpose of the weight vector of the step function. Δt represents the bias term of the step function, and exp(·) represents the exponential function.

[0039] In step 1.5, during the model training phase, the dataset processed in step 1.1 is used to construct the dual-branch model (TC-ResNet and TE-Liquid Time-Constant Networks) from steps 1.2 and 1.3 to fuse spatiotemporal features. Mixed-precision training is then employed to optimize the model and find the optimal combination of hyperparameters. Specifically, the steps include the following:

[0040] 5.1: Bayesian optimization is used to perform a global search for hyperparameters, with the multi-objective combined score as the optimization objective, to determine the optimal learning rate, weight decay and other parameters;

[0041] 5.2: Update model parameters using the Adaptive Moment Estimator with Weight Decay (AdamW) optimizer, with cosine annealing learning rate scheduling;

[0042] 5.3: Use SmoothL1Loss as the loss function, combined with gradient clipping and early stopping mechanisms to prevent overfitting.

[0043] In step 1.6, test set data is used to output surface wind speed prediction results through a trained surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] 1. By constructing a dual-branch architecture that integrates multi-source data fusion and spatiotemporal feature embedding, this method overcomes the technical bottlenecks of traditional numerical weather prediction (NWP) models, which struggle to capture micro-meteorological features in complex terrains, and single deep learning models, which cannot balance the correlation between spatiotemporal features. This approach significantly improves the accuracy and reliability of wind speed prediction under complex terrains, providing precise data support for wind energy resource assessment and energy system optimization.

[0046] 2. The first innovation of this invention lies in constructing an end-to-end spatiotemporal feature dual-branch architecture and introducing Temporal Embedding (TE) and Convolutional Block Attention (CBAM) mechanisms to solve the problem of insufficient representation of wind speed changes in complex terrain by traditional models. Traditional numerical weather prediction (NWP) models rely on global atmospheric dynamic equations, making it difficult to capture the local micro-meteorological effects induced by terrain such as mountains and coastlines. Furthermore, most single deep learning models lack the ability to collaboratively extract spatiotemporal features, resulting in large prediction errors for wind speed in complex terrain areas. This invention achieves dynamic joint modeling of spatiotemporal features through a dual-branch design. It innovatively introduces Temporal Embedding (TE) and Convolutional Block Attention (CBAM) mechanisms into the traditional Residual Network (ResNet) model to form a TC-ResNet architecture, dynamically focusing on the spatial correlation between terrain features such as altitude and slope and wind speed. Simultaneously, Temporal Embedding (TE) is introduced into Liquid Time-Constant Networks, converting time parameters such as season and diurnal time into differential equation modulation factors, thereby reducing prediction errors for periodic phenomena such as strong winter winds and afternoon valley winds. The overall model architecture adaptively allocates weights through a gating fusion mechanism, achieving deep coupling of spatiotemporal features. This results in a significant improvement in wind speed prediction accuracy in complex terrain areas compared to traditional methods.

[0047] 3. The second innovation of this invention lies in the introduction of a three-level feature fusion strategy for multi-source data, which solves the problems of noise interference and information redundancy in meteorological and topographic parameters. Traditional methods mostly use only meteorological numerical output results or simply splice topographic data, lacking correlation screening of multi-source features, making the model susceptible to interference from irrelevant parameters and resulting in insufficient feature utilization. This invention achieves efficient feature fusion through three-level processing: First, the CatBoost model is used to identify 8 core meteorological features and 3 topographic features from 27 meteorological parameters and 4 topographic parameters, effectively eliminating redundant information and increasing the model's focus on core features; second, the Convolutional Block Attention (CBAM) mechanism is used to modulate the selected features with channel-space-global three-level weights, improving the model's response speed to global patterns and enabling it to more sensitively capture wind speed change trends at macro scales; finally, temporal embedding (TE) technology maps 5 time parameters into feature vectors, deeply fusing them with spatiotemporal features, enabling the model to better understand and capture the seasonal fluctuations in wind speed, significantly enhancing the model's adaptability to wind speed changes at different time scales. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0049] Figure 1This invention provides a system framework diagram for a surface wind speed prediction method based on the fusion of multi-source data and spatiotemporal features. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] Example 1

[0052] See Figure 1 As shown in the figure, this embodiment provides a method for predicting surface wind speed based on the fusion of multi-source data and spatiotemporal features, specifically including the following steps:

[0053] (1-1) The meteorological parameters, topographic data of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3), and time parameters of the THORPEX Interactive Global Encyclopedia (TIGGE) project were preprocessed. Outliers were removed by spatiotemporal linear interpolation and 3σ rule and aligned to a unified grid. The gradient boosting decision tree (CatBoost) model was used to select a specific number of parameters that were most correlated with the wind speed data in the Japan 55-Year Reanalysis (JRA-55) from 27 meteorological parameters and 4 topographic parameters, respectively. The training set, validation set and test set were divided according to the year.

[0054] (1-2) The spatiotemporal features of wind speed data are extracted by using the ResNet module, incorporating temporal embedding (TE) technology, and combining it with the convolutional block attention mechanism (CBAM).

[0055] (1-3) By introducing the time embedding technique (TE) through the Liquid Time-Constant Networks module, the dynamic patterns of the time series of wind speed data are mined using the gating mechanism and meta-learning.

[0056] (1-4) By splicing the dual-branch features through a gating fusion mechanism, calculating the fusion weight, and then inputting it into the MLP, the surface wind speed is predicted.

[0057] (1-5) Bayesian optimization of hyperparameters is adopted, combined with the adaptive moment estimation (AdamW) optimizer with weight decay, cosine annealing learning rate and smooth L1 loss function to optimize model training;

[0058] (1-6) Use the trained surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features to perform short-term surface wind speed prediction on the test set and output the prediction results.

[0059] In step (1-1), the three data sources are preprocessed separately, specifically including the following steps:

[0060] (2-1) For the meteorological parameters (including 27 parameters such as 10-meter wind speed components u10 and v10, topography orog, soil moisture sm, etc.) of the THORPEX interactive global collection (TIGGE) project, the topographic data (ASTER GDEM V3) of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model (elevation, relief, slope, aspect) and time parameters (year, month, day, hour, season) were processed for missing values. Linear interpolation was used in the time dimension, and linear interpolation was first performed in the spatial dimension, followed by nearest neighbor interpolation.

[0061] (2-2) The 3σ rule was used to remove outliers that deviated from the mean by more than 3 times the standard deviation, and all data were uniformly interpolated to a 0.25°×0.25° grid to match the four time points of 06:00, 12:00, 18:00 and 24:00 every day;

[0062] (2-3) The importance of features was calculated using the gradient boosting decision tree (CatBoost) model. Through a two-stage feature selection method, eight key meteorological parameters were selected from the meteorological parameters of the THORPEX interactive global collection (TIGGE) project, and three key terrain parameters were selected from the terrain data of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3). Finally, the training set, validation set and test set were divided according to the year.

[0063] In step (1-2), a residual network (TC-ResNet) branch architecture incorporating temporal embedding (TE) and convolutional block attention (CBAM) is used to extract the spatiotemporal features of meteorological data from the dataset processed in step (1-1). Specifically, this includes the following steps:

[0064] (3-1) Spatial features are extracted using the ResNet architecture, and a Convolutional Block Attention (CBAM) mechanism is introduced. Channel attention is calculated by global average pooling to determine channel weights, as shown in the following formula:

[0065]

[0066] α=σ(W1·ReLU(W0·AvgPool(x)))

[0067] Where, AvgPool(x) cThis represents the result of global average pooling on the c-th channel of the input feature map x, where i represents the latitude index of the feature map (ranging from 1 to 48), j represents the longitude index of the feature map (ranging from 1 to 96), and x... c (i,j) represents the feature value located at grid point (i,j) in the c-th channel, α represents the generated channel attention weight matrix, σ represents the sigmoid activation function, W0 represents the dimension reduction convolution kernel, W1 represents the dimension increase convolution kernel, and ReLU represents the linear rectified activation function.

[0068] (3-2) Spatial attention extracts spatial cues through channel-dimensional average pooling and max pooling, as shown in the following formula:

[0069]

[0070] β=σ(Conv 7×7 (Concat(F avg ,F max )))

[0071] Among them, F avg (i,j) represents the result at grid point (i,j) after average pooling of the feature map along the channel dimension, 11 represents the total number of channels of the input feature, F max (i,j) represents the result at grid point (i,j) after max pooling the feature map along the channel dimension, β represents the generated spatial attention weight map, and Conv 7×7 This represents a 7×7 convolution operation that performs spatial feature extraction and dimensionality compression on the concatenated feature map. Concat(F) avg ,F max This indicates that the two results of channel pooling are concatenated along the channel dimension.

[0072] (3-3) A three-layer fully connected network is used to map the time parameters into a 56-dimensional time embedding vector, which enhances the ability to extract features of time dependencies. The formula is as follows:

[0073] TE(t)=W4·ReLU(W3·ReLU(W2·t))

[0074] Where TE(t) represents the output of the time embedding, t represents the set of time-related parameters of the input, W2 represents the weight matrix of the first fully connected network, W3 represents the weight matrix of the second fully connected network, and W4 represents the weight matrix of the third fully connected network.

[0075] In steps (1-3), the time-series dynamic patterns of wind speed data in the dataset processed in step (1-1) are mined using a branch architecture of TE-LiquidTime-Constant Networks, which incorporates temporal embedding (TE) technology, and utilizes gating mechanisms and meta-learning. This includes the following steps:

[0076] (4-1) The continuity of time series is modeled by ordinary differential equations based on liquid time-constant networks, and numerical integration is performed by the Runge-Kutta RK4 method.

[0077] (4-2) Based on the gating mechanism, which integrates time characteristics and input data, the formula is as follows:

[0078] g t =σ(W g [x t ;e′ t ]+b g ),f t =g t ⊙x t +(1-g t )⊙e′ t

[0079] Where g t W represents the gating weight at time step t. g The weight matrix represents the gating mechanism, used to perform a linear transformation on the concatenated features, x t The input data for time step t, e t ′ represents the temporal embedding feature after processing by a multilayer perceptron, b g The bias term representing the gating mechanism is a parameter learned during model training, used to adjust the baseline value of the linear transformation, f. t ⊙ represents the output of the gating mechanism;

[0080] (4-3) Time-varying weights and biases are generated from time features through a meta-learning mechanism to dynamically modulate the parameters of the ordinary differential equation (ODE) function;

[0081] (4-4) Adjust the fourth-order (RK4) integration step size based on time characteristics to enhance the modeling ability of non-stationary sequences. The formula is as follows:

[0082]

[0083] Wherein, Δt(e′) t ) represents the time-embedded feature e′ tThe calculated fourth-order (RK4) integral step size, where A and B represent hyperparameters used to adjust the overall scale and offset of the step size function, ensures that the step size is a reasonably positive value. b represents the transpose of the weight vector of the step function. Δt represents the bias term of the step function, and exp(·) represents the exponential function.

[0084] In steps (1-5), during the model training phase, the dataset processed in step (1-1) is used to fuse spatiotemporal features by constructing the dual-branch models (TC-ResNet and TE-Liquid Time-Constant Networks) from steps (1-2) and (1-3), and mixed-precision training is employed to optimize the model and find the optimal combination of hyperparameters. Specifically, the steps include the following:

[0085] (5-1) Bayesian optimization is used to perform a global search of hyperparameters, with the multi-objective combined score as the optimization objective, to determine the optimal learning rate, weight decay and other parameters.

[0086] (5-2) Update the model parameters using the weighted decay adaptive moment estimator (AdamW) optimizer, and use cosine annealing learning rate scheduling;

[0087] (5-3) Smooth L1 Loss is used as the loss function, and gradient clipping and early stopping mechanism are combined to prevent overfitting.

[0088] In step 1.6, test set data is used to output surface wind speed prediction results through a trained surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features.

[0089] On the same dataset, namely the eight meteorological parameters most correlated with wind speed data in the Japan 55-Year Reanalysis (JRA-55) selected from 27 meteorological parameters from the THORPEX Interactive Global Encyclopedia (TIGGE) project, the three topographic parameters most correlated with wind speed data in the Japan 55-Year Reanalysis (JRA-55) selected from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3) topographic data, and five time parameters, covering four time points each day from January 1, 2023 to December 31, 2023 (UTC 06:00, 12:00, 18:00, and 24:00). The surface wind speed prediction model based on the fusion of multi-source data and spatiotemporal features proposed in this invention was evaluated against existing Convolutional Long Short-Term Memory (ConvLSTM) models and European Centre for Medium-Range Weather Forecasts (ECMWF) models. The quality of the models was automatically evaluated using five performance indicators from the field of wind speed prediction research: root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), relative root mean square error (rRMSE), relative mean absolute error (rMAE), and percentage of samples with wind speed absolute error not exceeding 1 m / s (FA).

[0090] Table 1 Comparison of Results between the Method of the Present Invention and Other Methods

[0091]

[0092] Experiments show that the surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features proposed in this embodiment can achieve more accurate short-term surface wind speed prediction compared to the baseline method. Specifically, the model in this embodiment constructs a dual-branch architecture of TC-ResNet and TE-Liquid Time-Constant Networks, combined with the convolutional block attention mechanism (CBAM) and a three-level feature fusion strategy for multi-source data, effectively capturing the spatiotemporal dynamic features of wind speed under complex terrain. Furthermore, the model's perception of seasonal wind speed fluctuations is enhanced through temporal embedding (TE) technology, and deep feature integration is achieved using a gated fusion mechanism. This enables the model to exhibit excellent predictive capabilities under different geographical environments and meteorological conditions, surpassing these baseline methods. Specifically, for FA, the method in this embodiment can improve performance by at least 8.51%; for RMSE, the method can reduce error by at least 10.42%; for MAE, the method can reduce error by at least 11.11%; for rRMSE, the method can reduce error by at least 10.10%; for rMAE, the method can reduce error by at least 10.36%; and for R, the method can improve performance by at least 3.45%. These results demonstrate the competitiveness of the proposed method in this embodiment.

[0093] Example 2

[0094] To compare the performance of the model of this invention with existing surface wind speed prediction models in various years, a different dataset than that in Example 1 was selected for testing. Specifically, the dataset consisted of 8 meteorological parameters selected from 27 meteorological parameters in the THORPEX Interactive Global Encyclopedia (TIGGE) project that were most correlated with the wind speed data in the Japan 55-Year Reanalysis (JRA-55), 3 topographic parameters selected from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3) that were most correlated with the wind speed data in the Japan 55-Year Reanalysis (JRA-55), and 5 time parameters, covering four time points each day from January 1, 2024 to December 31, 2024 (UTC 06:00, 12:00, 18:00, and 24:00). The surface wind speed prediction model based on the fusion of multi-source data and spatiotemporal features proposed in this invention was evaluated against existing models such as Convolutional Long Short-Term Memory Network (ConvLSTM), Variational Mode Decomposition-Long Short-Term Memory Network (VMD-LSTM), and European Centre for Medium-Range Weather Forecasts (ECMWF). Five performance metrics from the field of wind speed prediction research (RMSE, MAE, R, rRMSE, rMAE, and FA) were used to automatically evaluate the quality of the models.

[0095] Table 2 Comparison of Results between the Method of the Present Invention and Other Methods

[0096]

[0097] Experiments show that the surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features proposed in this embodiment can achieve more accurate short-term surface wind speed prediction compared to the baseline method. Specifically, for FA, the method of this embodiment can improve performance by at least 10.78%; for RMSE, the method of this embodiment can reduce the error by at least 14.14%; for MAE, the method of this embodiment can reduce the error by at least 16.00%; for rRMSE, the method of this embodiment can reduce the error by at least 14.41%; for rMAE, the method of this embodiment can reduce the error by at least 15.51%; and for R, the method of this embodiment can improve performance by at least 5.81%. These results demonstrate the competitiveness of the method proposed in this embodiment.

[0098] Example 3

[0099] To compare the performance of the model of this invention with that of the European Centre for Medium-Range Weather Forecasts (ECMWF) model in various performance indicators for wind speed prediction at four time points each day (UTC 06:00, 12:00, 18:00, and 24:00), a portion of the dataset from Example 2 was selected for study. This dataset consists of eight meteorological parameters selected from 27 meteorological parameters in the THORPEX Interactive Global Encyclopedia (TIGGE) project that have the highest correlation with wind speed data in the Japan 55-Year Reanalysis (JRA-55); three topographic parameters selected from the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 (ASTER GDEM V3) topographic data that have the highest correlation with wind speed data in the Japan 55-Year Reanalysis (JRA-55); and five time parameters, covering four time points each day (UTC 06:00, 12:00, 18:00, and 24:00) from September 1, 2024 to September 30, 2024. September is a transitional season, and wind speeds exhibit a characteristic of fluctuating from the relative stability of summer to the multi-fluctuating winds of autumn, encompassing wind speed scenarios with varying intensities and patterns of change. Therefore, September 2024 was selected to evaluate the surface wind speed prediction model based on multi-source data and spatiotemporal feature fusion proposed in this invention, as well as the European Centre for Medium-Range Weather Forecasts (ECMWF) model. Five performance indicators from the field of wind speed prediction research (RMSE, MAE, R, rRMSE, rMAE, and FA) were used to automatically assess the model's quality.

[0100] Table 3 Comparison of results between the method of this invention and other methods

[0101]

[0102] Experiments show that the surface wind speed prediction model (MTRCL) based on the fusion of multi-source data and spatiotemporal features proposed in this embodiment can achieve more accurate short-term surface wind speed predictions compared to the European Centre for Medium-Range Weather Forecasts (ECMWF) model. Specifically, for FA, the method of this embodiment can improve performance by at least 32.13%; for RMSE, the method can reduce error by at least 29.66%; for MAE, the method can reduce error by at least 20.00%; for rRMSE, the method can reduce error by at least 34.65%; for rMAE, the method can reduce error by at least 36.17%; and for R, the method can improve performance by at least 18.67%. These results demonstrate the competitiveness of the proposed method.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting surface wind speed based on the fusion of multi-source data and spatiotemporal features, characterized in that, Includes the following steps: 1.1: The meteorological parameters, topographic data of ASTER GDEM V3 (Advanced Spaceborne Thermal Emission and Reflection Radiometer) from the THORPEX interactive global ensemble TIGGE project, and time parameters were preprocessed. Outliers were removed by spatiotemporal linear interpolation and 3σ rule and aligned to a unified grid. The gradient boosting decision tree CatBoost model was used to select a specific number of parameters from 27 meteorological parameters and 4 topographic parameters that were most correlated with the wind speed data in the 55-year reanalysis of Japan (JRA-55). The data were then divided into training, validation, and test sets according to the year. 1.2: Using the ResNet module of the residual network, incorporating temporal embedding technology, and combining the convolutional block attention mechanism CBAM, the spatiotemporal features of wind speed data are extracted; 1.3: By introducing time embedding technology through a liquid time constant network module, and utilizing gating mechanism and meta-learning, dynamic patterns of time series wind speed data are mined. 1.4: By splicing dual-branch features through a gating fusion mechanism, calculating the fusion weights, and then inputting them into the MLP, surface wind speed prediction is achieved; 1.5: Bayesian optimization of hyperparameters is adopted, combined with an adaptive moment estimator optimizer with weight decay, cosine annealing learning rate and smooth L1 loss function to optimize model training; 1.6: Use the trained surface wind speed prediction model based on the fusion of multi-source data and spatiotemporal features to perform short-term surface wind speed prediction on the test set and output the prediction results.

2. The surface wind speed prediction method based on multi-source data and spatiotemporal feature fusion according to claim 1, characterized in that, Step 1.1 includes the following steps: 2.1: Missing values ​​were processed for meteorological parameters, topographic data and time parameters of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3 from the THORPEX Interactive Global Collection project. Linear interpolation was used in the time dimension, and linear interpolation was first performed in the spatial dimension, followed by nearest neighbor interpolation. 2.2: The 3σ rule was used to remove outliers that deviated from the mean by more than 3 times the standard deviation, and all data were uniformly interpolated to a 0.25°×0.25° grid to match the four time points of 06:00, 12:00, 18:00 and 24:00 each day; 2.3: The gradient boosting decision tree model is used to calculate the importance of features. Through a two-stage feature selection method, eight key meteorological parameters are selected from the data of the THORPEX interactive global encyclopedia project, and three key terrain parameters are selected from the terrain data of the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global Digital Elevation Model Version 3. Finally, the training set, validation set and test set are divided according to the year.

3. The surface wind speed prediction method based on multi-source data and spatiotemporal feature fusion according to claim 1, characterized in that, Step 1.2 includes the following steps: 3.1: Spatial features are extracted using a residual network architecture, and a convolutional block attention mechanism is introduced. Channel attention is calculated using global average pooling to determine channel weights, as shown in the following formula: α=σ(W1·ReLU(W0·AvgPool(x))) Where, AvgPool(x) c This represents the result of global average pooling on the c-th channel of the input feature map x, where i represents the latitude index of the feature map (ranging from 1 to 48), j represents the longitude index of the feature map (ranging from 1 to 96), and x... c (i,j) represents the feature value located at grid point (i,j) in the c-th channel, α represents the generated channel attention weight matrix, σ represents the sigmoid activation function, W0 represents the dimension reduction convolution kernel, W1 represents the dimension increase convolution kernel, and ReLU represents the linear rectified activation function. 3.2: Spatial attention extracts spatial features through channel-dimensional average pooling and max pooling, as shown in the following formula: β=σ(Conv 7×7 (Concat(F avg ,F max ))) Among them, F avg (i,j) represents the result at grid point (i,j) after average pooling of the feature map along the channel dimension, 11 represents the total number of channels of the input feature, F max (i,j) represents the result at grid point (i,j) after max pooling the feature map along the channel dimension, β represents the generated spatial attention weight map, and Conv 7×7 This represents a 7×7 convolution operation that performs spatial feature extraction and dimensionality compression on the concatenated feature map. Concat(F) avg ,F max This indicates that the two results of channel pooling are concatenated along the channel dimension. 3.3: A three-layer fully connected network is used to map the time parameters into a 56-dimensional time embedding vector, enhancing the ability to extract features related to time dependencies. The formula is as follows: TE(t)=W4·ReLU(W3·ReLU(W2·t)) Where TE(t) represents the output of the time embedding, t represents the set of time-related parameters of the input, W2 represents the weight matrix of the first fully connected network, W3 represents the weight matrix of the second fully connected network, and W4 represents the weight matrix of the third fully connected network.

4. The surface wind speed prediction method based on multi-source data and spatiotemporal feature fusion according to claim 1, characterized in that, Step 1.3 includes the following steps: 4.1: Modeling the continuity of time series using ordinary differential equations based on liquid time constant networks, and performing numerical integration using the Runge-Kutta fourth-order method; 4.2: Based on the gating mechanism that integrates time characteristics and input data, the formula is as follows: g t =σ(W g [x t ;e′ t ]+b g ),f t =g t ⊙x t +(1-g t )⊙e′ t Among them, g t W represents the gating weight at time step t. g The weight matrix represents the gating mechanism, used to perform a linear transformation on the concatenated features, x t e′ represents the input data at time step t. t b represents the temporal embedding features after processing by a multilayer perceptron. g The bias term representing the gating mechanism is a parameter learned during model training, used to adjust the baseline value of the linear transformation, f. t ⊙ represents the output of the gating mechanism; 4.3: Time-varying weights and biases are generated from temporal features through a meta-learning mechanism to dynamically modulate the parameters of the differential equation ODE function; 4.4: Adjusting the fourth-order integral step size based on time characteristics enhances the ability to model non-stationary sequences. The formula is as follows: Wherein, Δt(e′) t ) represents the time-embedded feature e′ t The calculated fourth-order integral step size, where A and B represent hyperparameters used to adjust the overall scale and offset of the step size function, ensures that the step size is a reasonably positive value. b represents the transpose of the weight vector of the step function. Δt represents the bias term of the step function, and exp(·) represents the exponential function.

5. The surface wind speed prediction method based on the fusion of multi-source data and spatiotemporal features according to claim 1, characterized in that, Step 1.5 includes the following steps: 5.1: Bayesian optimization is used to perform a global search for hyperparameters, with the multi-objective combined score as the optimization objective, to determine the optimal learning rate and weight decay parameters; 5.2: Update model parameters using an adaptive moment estimator with weight decay and employ cosine annealing learning rate scheduling; 5.3: Use smoothed L1 as the loss function, combined with gradient clipping and early stopping mechanisms to prevent overfitting.

6. The surface wind speed prediction method based on multi-source data and spatiotemporal feature fusion according to claim 1, characterized in that, In step 1.6, test set data is used to output prediction results through a trained surface wind speed prediction model based on the fusion of multi-source data and spatiotemporal features.

Images