A target area wind speed and wind direction prediction method based on an embedded double-U network

Abstract

The application discloses a target area wind speed and wind direction prediction method based on an embedded double-U network, which comprises the following steps: obtaining a plurality of wind field data in a target area; pre-processing the wind field data by using a sliding window to obtain pre-processed wind field data; the step length and size of the sliding window are equal; constructing a standardization model to standardize the pre-processed wind field data to obtain standardized wind field data; inputting the standardized wind field data into an optimized embedded double-U network to obtain a target area wind speed and wind direction prediction result; the embedded double-U network comprises a non-parametric optimization space U-Net and a dilated Inception time U-Net. The method can be widely applied to the fields of wind energy resource assessment and wind farm layout design by integrating data preprocessing, feature standardization and embedded double-U network architecture, and realizes high-precision prediction of the wind speed and wind direction of a vast area wind field, thereby improving the scheduling and operation efficiency of a large regional power grid of a wind power center.

Description

Technical Field

[0001] This invention relates to the field of wind speed and wind direction prediction, and more specifically to a method for predicting wind speed and wind direction in a target area based on an embedded dual-U network. Background Technology

[0002] Currently, wind power, as a renewable energy source, possesses significant advantages due to its clean and sustainable attributes. However, it also has some drawbacks. The randomness of wind power output and the uncertainty of predicting wind energy supply impose certain limitations on its development. Because wind energy cannot be stored for future power generation, planning relies heavily on reliable short-term forecasts. Furthermore, the strong randomness and volatility of wind speed pose challenges to the safe and stable operation of wind farms. Relying solely on real-time wind speed data limits the responsiveness of the control system, leading to unstable wind power supply.

[0003] With the large-scale integration of wind power into the grid system, this instability can reduce economic benefits and even lead to "wind curtailment." Therefore, accurate short-term wind speed forecasts are crucial for the guidance, scheduling, operation, and maintenance of wind power.

[0004] In existing technologies, wind speed prediction mainly falls into two categories: physical model-based methods and data-driven methods. Physical model-based methods are effective for medium- to long-term predictions. For example, CFD-based microscale numerical simulation techniques for wind speed prediction establish a steady-state flow field based on inflow conditions, terrain, and surface roughness, but these require significant computational resources and time. Furthermore, this approach struggles to meet the timeliness requirements necessary for short-term predictions.

[0005] Data-driven methods infer future wind speed distributions by learning the evolution of historical wind speed data. While these methods offer higher accuracy, they require extensive historical data and statistical learning to identify underlying wind speed characteristics for prediction. However, as the number of prediction steps increases, the errors in these models tend to accumulate, making it difficult to capture the nonlinear characteristics of wind speed.

[0006] Machine learning-based methods are typical data-driven approaches, including artificial neural networks, SVMs, and the currently popular deep learning. These include using the Support Vector Regression (SVR) framework for short-term wind energy forecasting, highlighting SVR's predictive power for individual wind turbines at the microscale and its ability to predict the entire wind farm at a larger scale. Other methods involve designing and learning shallow neural structures with fewer levels of abstraction, such as feedforward ANNs, recurrent ANNs, RBF networks, ridge wave ANNs, and adaptive wavelet ANNs. These techniques can easily and quickly achieve predictions for simple, small-scale wind farm models. However, for historical data from large-scale wind farms, the more complex nonlinear spatiotemporal characteristics make it difficult for shallow machine learning models to easily capture evolutionary patterns.

[0007] Furthermore, while wind speed prediction networks based on improved empirical mode decomposition and GA-BP neural networks, and adaptive deep learning models capturing the dynamic characteristics of wind speed, are all single-point prediction methods, only able to predict the wind speed distribution at a single point. Moreover, because spatial wind speed distribution is influenced by multiple variables such as terrain, wind direction, and temperature, wind speeds at different locations exhibit temporal and spatial correlations. Considering only the wind speed at a single point and ignoring the correlations between different stations also limits the improvement of prediction accuracy.

[0008] It also includes a deep attention-based convolutional recurrent model based on K-Shape and enhanced memory, which integrates attention layers, convolutional neural networks, and recurrent neural networks (RNNs) to extract spatiotemporal latent representations and uses bidirectional gated recurrent units (Bi-GRUs) for temporal feature learning to achieve multi-point prediction. However, this technical solution, which converts spatial information into temporal information by flattening the spatial information into a one-dimensional vector, does not fully consider spatial relationships and cannot meet the needs of large-scale spatial wind speed prediction.

[0009] Therefore, how to design a target area wind speed and direction prediction method based on embedded dual-U network to respond to the future demand for precise prediction of wind power in a vast area and to help wind power ultra-short-term and short-term dispatch planning is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0010] In view of this, the present invention provides a method for predicting wind speed and direction in a target area based on an embedded dual-U network. By integrating multi-source wind field data and performing efficient preprocessing and standardization, an embedded dual-U network structure is designed, which significantly improves the accuracy and practicality of wind speed and direction prediction, and provides strong technical support for wind farm operation optimization, wind energy resource management, and wind power project planning.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A method for predicting wind speed and direction in a target area based on an embedded dual-U network includes:

[0013] Acquire multiple wind field data within the target area; the wind field data includes wind speed and wind direction data at multiple wind turbine hub heights;

[0014] The wind field data is processed using a sliding window to obtain wind field data after sliding window processing; the step size and size of the sliding window are equal.

[0015] A standardized model is constructed to standardize the wind field data after the sliding window processing, resulting in standardized wind field data.

[0016] The standardized wind field data is input into the optimized embedded dual-U network to obtain wind speed and wind direction prediction results within the target area; the embedded dual-U network includes: nonparametric optimization spatial U-Net and extended Inception temporal U-Net.

[0017] The standardized model is as follows:

[0018]

[0019] Where y” represents the preprocessed value, y' represents the original value of the wind field variable, and y mean y represents the average of the original values ​​of the wind field variables. std This represents the standard deviation of the original values ​​of the wind field variable.

[0020] Preferably, in the embedded dual-U network, the grid is divided based on the latitude and longitude information of the target area, and S is defined. t ∈R C×H×W For the wind speed and direction distribution at a certain moment, γ t ∈R T×C×H×W The wind speed and direction distribution over a specific period of time;

[0021] Using m historical wind field data points from different grids to predict n future wind field data points, this can be represented as:

[0022] (S m+1 ,S m+2 ,……,S m+n =EDU(S1,S2,……,S) m )

[0023] Wherein, EDU represents an embedded dual-U network, T represents the time scale, C represents the variable scale, H represents latitude and longitude, and W represents latitude and longitude width.

[0024] Preferably, the nonparametric optimized spatial U-Net includes: a spatial encoder and a spatial decoder;

[0025] The spatial encoder includes three nonparametric optimized downsampling convolutional blocks; the spatial decoder includes three nonparametric optimized upsampling convolutional blocks corresponding to the nonparametric optimized downsampling convolutional blocks.

[0026] The nonparametric optimized downsampling convolutional block includes: two 3*3 downsampling convolutional units and a nonparametric attention optimization unit; wherein each downsampling convolutional unit is equipped with a normalization and SiLU activation function;

[0027] The nonparametric optimized upsampling convolutional block includes: two 3*3 downsampling convolutional units, a nonparametric attention optimization unit, and a pixel shuffle upsampling unit; wherein each downsampling convolutional unit is equipped with a normalization and SiLU activation function.

[0028] Preferably, the non-parametric attention optimization unit introduces the SAM attention mechanism, and the energy formula for obtaining the target neuron k is:

[0029]

[0030] in, x i λ represents other neurons, N represents the number of neurons in each feature channel of the layer, and λ represents the regularization parameter.

[0031] And a scaling operator is used to refine the features:

[0032]

[0033] Where E represents all cross-channel and spatial dimensions

[0034] Preferably, the extended Inception time U-Net includes multiple layers of time encoding and decoding extended Inception units; each layer of time encoding and decoding extended Inception units is shortcut connected.

[0035] Preferably, the temporal encoding and decoding dilated Inception unit includes: an initial convolution and a multi-scale dilated convolution;

[0036] The initial convolution is:

[0037]

[0038] in, This represents the l-th layer features output by the spatial extraction module;

[0039] The multi-scale dilated convolution is:

[0040] FT '=ReLU(D1(t)+D2(t)+D3(t)+D4(t))

[0041] Where D1(t), D2(t), D3(t), and D4(t) represent convolutions with dilation rates of 1, 2, 3, and 4, respectively.

[0042] Preferably, the optimization process of the embedded dual-U network includes:

[0043] A hybrid loss function is constructed to optimize the embedded dual-U network. The hybrid loss function includes: global MSE loss, high-frequency loss, and wind direction loss.

[0044] The global loss is:

[0045]

[0046] Where y and These represent the predicted value and the actual value, respectively.

[0047] High-frequency information extraction is performed using bioorthogonal wavelets as a basis to obtain the high-frequency map in the wavelet domain. The edge loss is:

[0048]

[0049] Define the wind speeds in the u-direction and v-direction of the grid points as u g and v g The wind direction at the grid points is:

[0050] Ang = arctan(u g ,v g )

[0051] The actual and predicted values ​​of the grid point wind direction are as follows:

[0052] y Ang =arctan(u y ,v y )

[0053]

[0054] The angle loss is:

[0055]

[0056] The mixed loss is:

[0057] L=ωL Glo +(1-ω)(L H +L Ang )

[0058] Where ω represents the equilibrium parameter.

[0059] Preferably, the optimization process of the embedded dual-U network includes: using root mean square error (RMSE), mean absolute error (MAE), accuracy (ACC), and ANG. a The model optimizes the embedded dual-U network.

[0060] Preferably, the optimization process of the embedded dual-U network includes: using ANG a The model optimizes the embedded dual-U network:

[0061] The wind direction angle range predicted by the model from 0° to 360° is divided into α angles; the angle is 360° / α.

[0062] Calculate the difference between all predicted angle values ​​and actual angle values ​​in the grid.

[0063] Compare all differences with 360° / α to obtain the number of differences that are less than 360° / α;

[0064] The proportion of the number of differences less than 360° / α in the total number of differences is calculated, and the accuracy of the wind direction prediction is determined based on a preset threshold and a preset level.

[0065] As can be seen from the above technical solutions, compared with the prior art, the technical solutions of the present invention have the following advantages:

[0066] Beneficial effects:

[0067] 1. By acquiring multiple wind field data within the target area, including wind speed and direction information at different turbine hub heights, this scheme can integrate multi-scale and multi-height wind field characteristics, improving the model's understanding of complex wind field environments and its prediction accuracy.

[0068] 2. Using the sliding window technique to preprocess the raw wind field data can flexibly capture the time series characteristics of the wind field data, maintain the continuity and spatiotemporal correlation of the data, and provide stable and reliable input for subsequent analysis.

[0069] 3. Standardize the preprocessed wind field data by constructing a standardized model. This eliminates the influence of dimensions, makes the model sensitive to each feature in a consistent manner, avoids optimization bias caused by different feature scales, and improves the model's generalization ability.

[0070] 4. By combining nonparametric optimization spatial U-Net and extended Inception temporal U-Net, the embedded dual-U network can perform comprehensive and in-depth learning in both spatial and temporal dimensions, improving the accuracy and robustness of predictions. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0072] Figure 1 This is a flowchart of the method for predicting wind speed and direction in a target area provided in an embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the main frame of the embedded dual-U network provided in an embodiment of the present invention;

[0074] Figure 3 A schematic diagram of the extended Inception temporal U-Net framework provided in an embodiment of the present invention;

[0075] Figure 4 This is a schematic diagram of RMSE results based on different algorithms provided in embodiments of the present invention;

[0076] Figure 5 This is a schematic diagram of MAE results based on different algorithms provided in the embodiments of the present invention;

[0077] Figure 6 This is a schematic diagram of ACC results based on different algorithms provided in embodiments of the present invention;

[0078] Figure 7 This is a wind rose diagram showing the actual and predicted wind speed and direction at location A, provided in an embodiment of the present invention.

[0079] Figure 8 This is a wind rose diagram showing the actual and predicted wind speed and direction at location B, provided in an embodiment of the present invention.

[0080] Figure 9 Wind rose diagrams of actual and predicted wind speed and direction at location C provided in this embodiment of the invention;

[0081] Figure 10 This is a schematic diagram showing the comparison of RMSE results at different times based on an embedded dual-U network, provided in an embodiment of the present invention. Detailed Implementation

[0082] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Example 1:

[0084] like Figure 1 As shown, this embodiment provides a method for predicting wind speed and direction in a target area based on an embedded dual-U network, including:

[0085] Acquire multiple wind field data within the target area; the wind field data includes wind speed and wind direction data at multiple wind turbine hub heights;

[0086] The wind field data is processed using a sliding window to obtain wind field data after sliding window processing; the step size and size of the sliding window are equal.

[0087] A standardized model is constructed to standardize the wind field data after the sliding window processing, resulting in standardized wind field data.

[0088] The standardized wind field data is input into the optimized embedded dual-U network to obtain wind speed and wind direction prediction results within the target area; the embedded dual-U network includes: nonparametric optimization spatial U-Net and extended Inception temporal U-Net.

[0089] The aforementioned method for predicting wind speed and direction in the target area performs efficient preprocessing and standardization of wind field data, and improves the accuracy and practicality of wind speed and direction prediction based on an embedded dual-U network. It can meet the future demand for fine prediction of wind power generation in a vast area and improve the scheduling and operation efficiency of wind power centers for large regional power grids.

[0090] The following provides a further detailed explanation of the methods for predicting wind speed and direction in the aforementioned target areas:

[0091] In this embodiment, wind farm data is sourced from ground-based meteorological stations, upper-air sounding equipment (such as radiosondes), remote sensing satellites, and sensors specifically deployed at the wind farm (such as anemometers and wind vanes). Because hub heights vary in different areas within the target region, data collection covers different altitude levels.

[0092] The wind field data is preprocessed using a sliding window, with a size of W and a step size of W. The sliding window covers the entire data sequence sequentially, ensuring no data points are missed, and generates a new data segment for analysis with each movement.

[0093] A standardized model is constructed to standardize the wind field data after sliding window processing. The standardized model is as follows:

[0094]

[0095] Where y” represents the preprocessed value, y' represents the original value of the wind field variable, and y mean y represents the average of the original values ​​of the wind field variables. std This represents the standard deviation of the original values ​​of the wind field variable.

[0096] In this embodiment, the construction and optimization process of the embedded dual-U network includes:

[0097] Dataset and Dataset Processing: The dataset used for training the embedded dual-U network is wind field data from a certain region. Taking a certain city in the region as an example, the height of its wind turbines is about 100 meters. Therefore, the meridional and zonal winds at a height of 100 meters are selected for wind speed and direction prediction, and the time range and dataset division are determined.

[0098] The dataset is processed using a sliding window, and the system is set to predict wind speed data for the next 24 hours based on 24 hours of historical data.

[0099] Furthermore, because the data scales are not within the same range, this can affect the model's training results. To obtain data more suitable for model training, the data is standardized.

[0100] The embedded dual-U network enables the prediction of wind speed and direction in large-scale wind fields, involving a four-dimensional problem encompassing time, space, and variables. The wind field is divided into grids based on spatial information (longitude and latitude), with each grid serving as an observation point for multivariate data, and each cycle containing a set of multivariate values.

[0101] Define S t ∈R C×H×W For the wind speed and direction distribution at a certain moment, γ t ∈R T×C×H×W The wind speed and direction distribution over a specific period of time;

[0102] Using m historical wind field data points from different grids to predict n future wind field data points, this can be represented as:

[0103] (S m+1 ,S m+2 ,……,S m+n =EDU(S1,S2,……,S) m )

[0104] Wherein, EDU represents an embedded dual-U network, T represents the time scale, C represents the variable scale, H represents latitude and longitude, and W represents latitude and longitude width.

[0105] In this embodiment, as Figure 2 As shown, the embedded dual-U network includes a nonparametric optimization space U-Net and an extended Inception time U-Net; its data processing procedure is as follows:

[0106] The input data is F, representing the regional wind speed and direction distribution over historical time periods u. The input data is then reshaped to obtain Fs. Fs is input into the spatial encoder of the nonparametric optimized spatial Unet for spatial feature extraction, sequentially outputting multi-scale spatial features. These multi-scale features are then reshaped. The corresponding extended Inception time Unet is input for time dependency extraction to capture the temporal evolution of the wind field. Simultaneously, the nonparametric optimized spatial Unet decoder is also input for spatial feature decoding. Finally, the multi-scale features are integrated to obtain the final output: the wind speed and direction distribution at a future time.

[0107] The nonparametric optimization spatial U-Net consists of a spatial encoder and a spatial decoder; the spatial encoder is responsible for extracting the spatial features of the wind field, and the spatial decoder is responsible for reconstructing the wind field at future moments.

[0108] The spatial encoder comprises three nonparametric optimized downsampling convolutional blocks; when the input data size is R... 2×H×W When the feature scales obtained are R, 16×H / 2×W / 2 R 32×H / 4×W / 4 R 64×H / 8×W / 8 Among them, R 16×h / 2×W / 2 These are shallow features with more texture information, responsible for extracting and reconstructing wind field texture information; R 64×h / 8×W / 8 As a high-level feature, it possesses more advanced semantic features and can represent the spatial features of wind fields over a larger area, which can be used to extract and reconstruct wind field semantic information.

[0109] The spatial decoder includes three nonparametric optimized upsampling convolutional blocks corresponding to the nonparametric optimized downsampling convolutional blocks;

[0110] The nonparametric optimized downsampling convolutional block includes two 3*3 downsampling convolutional units and a nonparametric attention optimization unit; each downsampling convolutional unit is equipped with a normalization and SiLU activation function;

[0111] The nonparametric optimized upsampling convolutional block includes: two 3*3 downsampling convolutional units, a nonparametric attention optimization unit, and a pixel shuffle upsampling unit; each downsampling convolutional unit is equipped with a normalization and SiLU activation function.

[0112] By upsampling features at different scales, features at different scales are effectively fused, ensuring that the model decoder can acquire semantic and texture features after time evolution, thereby ensuring the reconstruction quality of future wind fields.

[0113] Furthermore, the spatial characteristics of wind fields are more complex than those of images, and traditional convolution cannot fully extract key features. Wind field variations are complex, influenced by topography and other factors, resulting in significant fluctuations in wind speed across different locations, manifesting spatially as rich textures and edge features. Attention is particularly effective in extracting these textures and edge features. Therefore, we introduce nonparametric optimized attention into our convolutional block design to enhance the spatial extraction capabilities of the convolutional blocks.

[0114] The input channels contain u and v, and their spatial distribution is rugged and undulating. Therefore, using SAM to explore the joint attention weight distribution of the channel-space is more suitable for feature extraction and reconstruction of wind fields. SAM obtains the attention weight of a neuron by considering the importance of that neuron in the feature. We measure importance by measuring the linear separability of a target neuron from other neurons.

[0115] The nonparametric attention optimization unit introduces the SAM attention mechanism, and the energy formula for obtaining the target neuron k is:

[0116]

[0117] in, x i λ represents other neurons, N represents the number of neurons in each feature channel of the layer, and λ represents the regularization parameter.

[0118] And a scaling operator is used to refine the features:

[0119]

[0120] Where E represents all cross-channel and spatial dimensions

[0121] With the assistance of this non-parametric attention mechanism, our convolutional blocks can capture important features in the feature space and adjust the focus areas of the wind field feature space in real time. The attention mechanism allows the model to assign personalized weights to different spatial locations. This means that the model can more flexibly treat the contribution of different regions to the overall prediction, improving the model's adaptability to spatial heterogeneity.

[0122] like Figure 3 As shown, the extended Inception temporal U-Net includes multiple layers of temporal encoding and decoding extended Inception units; each layer of temporal encoding and decoding extended Inception units is shortcut-connected.

[0123] The temporal encoding and decoding dilated Inception unit includes: initial convolution and multi-scale dilated convolution;

[0124] The initial convolution is:

[0125]

[0126] in, This represents the l-th layer features output by the spatial extraction module;

[0127] The multi-scale dilated convolution is:

[0128] F T '=ReLU(D1(t)+D2(t)+D3(t)+D4(t))

[0129] Where D1(t), D2(t), D3(t), and D4(t) represent convolutions with dilation rates of 1, 2, 3, and 4, respectively.

[0130] By employing multi-layer convolutions, the temporal evolution of wind fields can be captured. Utilizing multi-scale convolutional arrangements can address the problem of unstable spatial locations of important features. Furthermore, dilated convolutions have a wider receptive field than ordinary convolutions; combining ordinary and dilated convolutions allows our model to extract all-to-local effective information from temporally dynamic features.

[0131] The weather data obtained in this embodiment is mostly 0.25*0.25 grid data, with an actual distance of approximately 27.5 kilometers near the equator. This increases the likelihood of sudden weather changes. In particular, the spatiotemporal fields such as wind fields and precipitation fluctuate dramatically, often resulting in areas of spatial variation. These points pose significant challenges to our predictions; simply minimizing the MSE loss function is insufficient to capture the high-frequency information response to sudden changes and produce satisfactory results. Furthermore, due to the inherent variability of wind direction, the global MSE loss is insufficient to specifically capture wind direction. To adapt to the unpredictable nature of wind fields and enable our model to better capture the spatiotemporal patterns of wind speed and direction, a hybrid loss function is constructed to optimize the embedded dual-U network, including: global MSE loss, high-frequency loss, and wind direction loss.

[0132] The global loss is:

[0133]

[0134] Where y and These represent the predicted value and the actual value, respectively.

[0135] High-frequency information extraction is performed using bioorthogonal wavelets as a basis to obtain the high-frequency map in the wavelet domain. The edge loss is:

[0136]

[0137] Define the wind speeds in the u-direction and v-direction of the grid points as u g and v g The wind direction at the grid points is:

[0138] Ang = arctan(u g ,v g )

[0139] The actual and predicted values ​​of the grid point wind direction are as follows:

[0140] y Ang =arctan(u y ,v y )

[0141]

[0142] The angle loss is:

[0143]

[0144] The mixed loss is:

[0145] L=ωL Glo +(1-ω)(L H +L Ang )

[0146] Where ω represents the equilibrium parameter.

[0147] In this embodiment, the root mean square error (RMSE), mean absolute error (MAE), accuracy (ACC), and ANG are used. a The model optimizes the embedded dual-U network. To reflect the location differences at different latitudes, a latitude-weighted evaluation metric is used.

[0148] Using ANG a The model optimizes the embedded dual-U network, including:

[0149] The wind direction angle range predicted by the model from 0° to 360° is divided into α angles; the angle is 360° / α.

[0150] Calculate the difference between all predicted angle values ​​and actual angle values ​​in the grid.

[0151] Compare all differences with 360° / α to obtain the number of differences that are less than 360° / α;

[0152] The proportion of the number of differences less than 360° / α in the total number of differences is calculated, and the accuracy of the wind direction prediction is determined based on a preset threshold and a preset level.

[0153] The formula is:

[0154]

[0155] For a given HW grid point at a given moment, we can obtain a total of HW wind speed angles. We divide the 360-degree angle into equal parts, typically with values ​​of 4, 8, 16, and 32.

[0156] Example 2:

[0157] like Figures 4-6 As shown, in order to prove the accuracy of the network model in this technical solution in wind speed prediction, the network model in this technical solution is compared with the classic prediction method. In the comparison process, historical 24 data is used to predict the wind speed and wind direction distribution for the next 24 hours.

[0158] The following table is based on ANG a The model evaluation compares different network models and wind direction, with a total of 64*80 points. The percentage of each point that differs from the true value by less than the following preset threshold is considered.

[0159]

[0160]

[0161] Based on the RMSE, MAE, and ACC results of different algorithms in predicting wind speed for the next 24 days, and based on ANG a The evaluation results of the model show that the embedded dual-U network (EDU) in this technical solution has outperformed other models in terms of the accuracy of generating fine grid wind field data in a certain region.

[0162] Its high accuracy in short-term forecasting is crucial for the real-time scheduling and safe operation of wind farms. It can effectively help wind power centers plan power production and grid connection strategies in advance, and reduce energy waste or supply shortages caused by wind speed forecasting errors.

[0163] In this embodiment, as Figures 7-9 As shown, in order to verify the accuracy of the wind direction prediction of this technical solution for the whole year, we use a wind rose diagram to illustrate the wind direction prediction of three locations A, B and C in a certain region.

[0164] Based on the actual and predicted wind rose diagrams for locations A, B, and C, the annual wind speed distribution shows that the 0-3 m / s range is relatively rare. At location A, the dominant wind directions are northeast and southwest, with wind speeds mostly between 3-12 m / s. At location B, the dominant wind directions are northeast and westerly, with wind speeds also mostly between 3-12 m / s, but northeasterly winds are predominantly stronger. At location C, northeasterly winds are dominant, with slightly weaker southwesterly winds, and wind speeds are also slightly weaker than at locations A and B. Comparing the predicted and actual wind field distributions, it can be seen that this technical solution can effectively construct wind fields using fine-grained grids, and it has a good generation effect on wind speed and direction in key areas.

[0165] In this embodiment, as Figure 10 As shown, in order to measure the robustness of historical 24-hour data for learning and predicting different time periods in the future, historical 24-hour data was used to train and predict five time periods of 1, 3, 6, 12, and 24 hours in the future, resulting in five prediction models.

[0166] When predicting wind farm data for only the next 1 hour, the RMSE of the embedded dual-U network in this technical solution is only 0.38. As the prediction time increases, the model's prediction performance gradually decreases. However, for models used to predict different historical periods, the prediction performance is not significantly different. For example, the model predicting wind farm data for the next 24 hours performs almost identically in generating 12-hour wind farm data as the model predicting only the next 12 hours. This demonstrates that the embedded dual-U network in this technical solution exhibits good robustness when predicting wind farm data across different time periods, providing strong support for the refined management and operation of local wind farms.

[0167] In this embodiment, to demonstrate the model's timeliness performance, an efficiency comparison of various classic spatiotemporal prediction models regarding wind speed and direction prediction was also conducted, as shown in the table below:

[0168] ConvLSTM 32 111 PredRNN 44 16 E3D-LSTM 193 39 SimVP 24 81 STDGN 110 117 EDU 26 16

[0169] When predicting wind speed and direction distribution over the next 24 hours, E3D-LSTM took the longest time, requiring 193ms. This is because 3D convolution has more parameters, making computation slower. ConvLSTM took 32ms to predict a single task, but its GFLOPS reached 111G. Although STDGN uses a separate attention mechanism to greatly reduce the number of parameters in self-attention, its floating-point computation still reached 117G, and its inference time was also relatively long at 110ms.

[0170] The embedded dual-U network (EDU) in this technical solution uses pure 2D convolution to construct a spatiotemporal prediction model for wind farms, requiring only 16G of floating-point computation and a prediction time of only 26ms. These results demonstrate that the model exhibits superior efficiency, achieving better inference accuracy with less computation and faster inference time. It not only reduces hardware requirements but also paves the way for real-time applications on edge devices, such as direct integration into wind farm monitoring systems to enable instant wind speed and direction prediction and response.

[0171] Furthermore, the embedded dual-U network (EDU) in this technical solution possesses high portability, enabling rapid adaptation to wind speed and direction forecasting tasks in new regions, reducing the significant computational resources and time required for repeated training. It is of great value for rapid deployment to wind farms on a larger scale, greatly improving the ease of use and cost-effectiveness of the application. Combined with its strong versatility and flexibility, it is expected to play a crucial role in broader weather forecasting and renewable energy management.

[0172] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0173] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting wind speed and direction in a target area based on an embedded dual-U network, characterized in that, include: Acquire multiple wind field data within the target area; the wind field data includes wind speed and wind direction data at multiple wind turbine hub heights; The wind field data is processed using a sliding window to obtain wind field data after sliding window processing; the step size and size of the sliding window are equal. A standardized model is constructed to standardize the wind field data after the sliding window processing, resulting in standardized wind field data. The standardized wind field data is input into the optimized embedded dual-U network to obtain wind speed and wind direction prediction results within the target area. The embedded dual-U network includes: a nonparametric optimized spatial U-Net and an extended Inception temporal U-Net; The nonparametric optimized spatial U-Net includes: a spatial encoder and a spatial decoder; The spatial encoder includes three nonparametric optimized downsampling convolutional blocks; the spatial decoder includes three nonparametric optimized upsampling convolutional blocks corresponding to the nonparametric optimized downsampling convolutional blocks; the nonparametric optimized downsampling convolutional block includes: two 3*3 downsampling convolutional units and a nonparametric attention optimization unit; wherein each downsampling convolutional unit is equipped with a normalization and SiLU activation function; the nonparametric optimized upsampling convolutional block includes: two 3*3 downsampling convolutional units, a nonparametric attention optimization unit, and a pixel shuffle upsampling unit; wherein each downsampling convolutional unit is equipped with a normalization and SiLU activation function; The extended Inception temporal U-Net includes multiple layers of temporal encoding and decoding extended Inception units; each layer of temporal encoding and decoding extended Inception units is shortcut-connected. The embedded dual-U network's data processing includes: input data F, representing the regional wind speed and direction distribution over historical u time periods; and reshaping the input data F to obtain... ;Will Spatial features are extracted from the spatial encoder portion of the nonparametric optimized spatial Unet and multi-scale spatial features are output sequentially. The multi-scale features are then reshaped. Temporal dependencies are extracted from the corresponding extended Inception time Unet to capture the temporal evolution of the wind field. Simultaneously, the spatial features are decoded from the nonparametric optimized spatial Unet decoder. Finally, the multi-scale features are integrated to obtain the final output of the wind speed and direction distribution at a future time.

2. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The standardized model is as follows: in, Indicates the preprocessed value. Represents the original values ​​of the wind field variables. This represents the average of the original values ​​of the wind field variables. This represents the standard deviation of the original values ​​of the wind field variable.

3. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, In the embedded dual-U network, the grid is divided based on the latitude and longitude information of the target area, and the following definition is made. The wind speed and direction distribution at a certain moment. The wind speed and direction distribution over a specific period of time; Using m historical wind field data points from different grids to predict n future wind field data points, this can be represented as: Wherein, EDU represents an embedded dual-U network, T represents the time scale, C represents the variable scale, H represents latitude and longitude, and W represents latitude and longitude width.

4. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The non-parametric attention optimization unit introduces the SAM attention mechanism, and the energy formula for obtaining the target neuron k is: in, , ; Other neurons represent the number of neurons in each feature channel of the layer. Indicates the regularization term parameter; And a scaling operator is used to refine the features: Where E represents all cross-channel and spatial dimensions .

5. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The temporal encoding and decoding dilated Inception unit includes: initial convolution and multi-scale dilated convolution; The initial convolution is: in, This represents the l-th layer features output by the spatial extraction module; The multi-scale dilated convolution is: in, , , and These represent convolutions with dilation rates of 1, 2, 3, and 4, respectively.

6. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The optimization process of the embedded dual-U network includes: A hybrid loss function is constructed to optimize the embedded dual-U network. The hybrid loss function includes: global MSE loss, high-frequency loss, and wind direction loss. The global loss is: in, and These represent the predicted value and the actual value, respectively. High-frequency information extraction is performed using bioorthogonal wavelets as a basis to obtain the high-frequency map in the wavelet domain. The edge loss is: Define the wind speeds in the u-direction and v-direction at the grid points as follows: and The wind direction at the grid points is: The actual and predicted values ​​of the grid point wind direction are as follows: The angle loss is: The mixed loss is: in, This represents the equilibrium parameter.

7. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The optimization process of the embedded dual-U network includes: using root mean square error (RMSE), mean absolute error (MAE), accuracy (ACC), and... The model optimizes the embedded dual-U network.

8. The method for predicting wind speed and direction in a target area based on an embedded dual-U network according to claim 1, characterized in that, The optimization process of the embedded dual-U network includes: employing The model optimizes the embedded dual-U network: Range of wind direction angles predicted by the model Divided into equal parts An angle; the angle is ; Calculate the difference between all predicted angle values ​​and actual angle values ​​in the grid. sum all differences Compare and obtain the difference less than The number of differences; Calculate the difference to be less than The percentage of the number of differences in the total number of differences is used to determine the accuracy of wind direction prediction based on preset thresholds and preset levels.

Images