Wind speed spatiotemporal synchronous prediction method based on CAEL-UNet

Abstract

This invention discloses a spatiotemporal synchronous wind speed prediction method based on CAEL-UNet. The method involves acquiring long-term global satellite remote sensing data related to meteorological and environmental elements from publicly available data sources, and extracting two sets of wind speed data: easterly wind component data and northerly wind component data. The easterly and northerly wind component data are preprocessed, normalized, and then stacked together at the same time and space to form a spatiotemporal data cube. The dataset is then divided into training and testing sets according to time order. This invention significantly reduces computational resource consumption. CAEL-UNet uses a convolutional autoencoder to perform nonlinear dimensionality reduction on the high-dimensional input data, compressing the data dimension while preserving key spatiotemporal features, greatly reducing the memory usage and training time for model training, and meeting the timeliness requirements of real-time wind farm scheduling.

Description

Technical Field

[0001] This invention relates to the field of meteorological feature data prediction technology, specifically to a spatiotemporal synchronous prediction method for wind speed based on CAEL-UNet. Background Technology

[0002] As the global energy structure transitions towards a clean and low-carbon model, wind energy, as a core component of renewable energy, plays a strategic role in achieving the "dual carbon" goal through its efficient development and utilization. Wind speed, a key parameter for wind power generation, wind farm operation and maintenance, and grid dispatch, directly determines the power generation efficiency and grid connection stability of wind farms—accurate wind speed forecasting can effectively reduce wind power prediction errors, significantly reduce grid peak-shaving pressure, and improve wind energy absorption. However, wind speed data exhibits strong nonlinearity, spatiotemporal dynamic coupling, and random fluctuations, and is influenced by factors such as topography and meteorological systems. Traditional forecasting methods struggle to meet the accuracy and timeliness requirements of engineering applications.

[0003] Wind speed forecasting technology has gone through three stages of development: early numerical weather prediction methods based on physical mechanisms, which achieved prediction by solving atmospheric dynamic equations, but suffered from high computational complexity and significant short-term forecast lag; mid-term statistical learning methods such as ARIMA and Support Vector Machine (SVM), which built prediction models by mining statistical patterns in historical data, reduced computational costs, but struggled to handle the nonlinear and time-varying characteristics of wind speed data; in recent years, deep learning technology has become a research hotspot, and methods based on recurrent neural networks (RNN) and long short-term memory networks (LSTM) can effectively capture temporal dependencies, while the fusion model of LSTM and UNet (such as LSTM-UNet) further combines the multi-scale spatial feature extraction capability of UNet with the temporal dynamic modeling capability of LSTM, showing superior performance to traditional methods in spatiotemporal wind speed forecasting.

[0004] Despite advancements in current technology, wind speed forecasting still faces three core challenges: First, the computational cost of high-dimensional data is prohibitive. To improve prediction accuracy, LSTM-UNet models often require high-resolution spatiotemporal data (such as minute-by-minute wind speeds from multiple stations and meteorological grid data), leading to a dramatic increase in the dimensionality of input features. For example, with 256×256 meteorological grids and 24-hour time-series data, the model's feature parameters reach millions, requiring significant GPU memory (e.g., over 20GB on a professional GPU), and a single training run takes over 30 minutes, making it difficult to meet the timeliness requirements of real-time wind farm scheduling. Second, feature redundancy leads to low model efficiency. Actual wind speed data contains substantial spatial redundancy (e.g., correlations between wind speeds from adjacent wind farms exceeding 0.8) and temporal redundancy (e.g., minimal wind speed changes during stable periods). Existing LSTM-UNet models directly from the raw high-dimensional data without effectively compressing redundant features, increasing computational burden and potentially introducing noise interference, thus affecting prediction accuracy. Third, balancing dimensionality reduction and feature preservation is difficult. Some studies use traditional linear dimensionality reduction methods such as PCA to compress data, but these methods cannot capture the nonlinear spatiotemporal correlation of wind speed data and are prone to losing key dynamic features (such as wind speed gradient changes before gusts arrive), resulting in a significant decrease in the prediction performance of LSTM-UNet after dimensionality reduction. Summary of the Invention

[0005] The purpose of this invention is to provide a spatiotemporal synchronous prediction method for wind speed based on CAEL-UNet, so as to solve the technical problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Step 1: Obtain long-term global satellite remote sensing data related to meteorological and environmental elements from publicly available data sources, and extract two sets of wind speed data: easterly wind component data and northerly wind component data.

[0008] Step 2: Preprocess the easterly wind component data and the northerly wind component data, normalize the preprocessed data, stack the easterly wind component data and the northerly wind component data of the same time and space together to form a spatiotemporal data cube, and then divide the dataset into training set and test set according to time order.

[0009] Step 3: Integrate the convolutional autoencoder into the traditional LSTM-UNet model to construct the CAEL-UNet model. This model adopts a cascaded architecture: the front end is the encoder part of the convolutional autoencoder, the middle is the LSTM-UNet spatiotemporal prediction module, and the back end is the decoder part of the convolutional autoencoder.

[0010] Step 4: Input the training set data into the CAEL-UNet model for learning and training, and use the mean squared error as the loss function for training the model;

[0011] Step 5: Input the test set data into the trained CAEL-UNet model for prediction.

[0012] Furthermore, the specific steps of step S2 are as follows:

[0013] Step 21: Preprocess the outliers and missing values ​​of the easterly wind component data and the northerly wind component data using the spatial neighborhood filling method. The calculation formula is as follows:

[0014]

[0015] In the formula, M represents the grid data of the easterly and northerly wind components, with dimensions of rows × columns. The values ​​are the filled results. U / D / L / R are the valid values ​​of the upper / lower / left / right neighbors of the outlier and missing value locations, and K is the number of valid neighbors, 1≤K≤4;

[0016] Step 22: Normalize the preprocessed data. The calculation formula is as follows:

[0017]

[0018] In the formula, The minimum value among all samples of a single wind speed data point. Then and on the contrary, For a single sample value to be normalized, These are the normalized sample values. ∈[0,1];

[0019] Step 23: Stack the normalized data vertically to package the wind speed data into three-dimensional data. Each three-dimensional data includes the easterly wind and northerly wind at the same time. Finally, divide the preprocessed data according to the time scale.

[0020] Furthermore, in step 3, the convolutional autoencoder uses a symmetric encoder-decoder architecture for feature compression and reconstruction. The encoder part uses three consecutive convolutional blocks, including 3×3 convolution, LeakyReLU activation, batch normalization, and max pooling operations. The convolutional autoencoder introduces Dropout regularization after each convolutional layer.

[0021] Furthermore, the implementation of the convolutional autoencoder in step 3 is as follows: the encoder part of the convolutional autoencoder extracts and downsamples features from each frame of the wind speed data to generate a series of multi-scale spatial feature maps. Then, the obtained multi-scale spatial feature map sequence is input into the LSTM-UNet model. The UNet part is responsible for extracting spatial features from the input image. Through convolution and pooling operations, features of different scales and levels in the image are extracted. The LSTM part learns the temporal dependencies between different images and outputs a feature sequence that integrates spatiotemporal context information. Finally, the decoder part of the convolutional autoencoder upsamples and reconstructs features from the feature sequence that integrates spatiotemporal context information to output the final prediction result.

[0022] Furthermore, the dropout ratio of the encoder section is 0.1-0.2, and the decoder section is symmetrically configured.

[0023] Furthermore, the CAEL-UNet model employs a two-stage training strategy: first, a convolutional autoencoder is pre-trained separately, and then fine-tuned jointly with LSTM-UNet.

[0024] Beneficial effects:

[0025] 1. The present invention significantly reduces the consumption of computing resources. CAEL-UNet performs nonlinear dimensionality reduction on high-dimensional input data through a convolutional autoencoder, compressing the data dimension while retaining key spatiotemporal features, greatly reducing the memory usage and training time of the model, and meeting the timeliness requirements of real-time scheduling of wind farms.

[0026] 2. The present invention ensures stable prediction accuracy. The convolutional autoencoder achieves effective feature compression through nonlinear mapping, avoiding the loss of key spatiotemporal features caused by traditional linear dimensionality reduction. Combined with the spatiotemporal modeling capabilities of LSTM-UNet, the wind speed prediction performance is better than existing dimensionality reduction schemes. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the spatiotemporal synchronous prediction method for wind speed based on CAEL-UNet of the present invention;

[0029] Figure 2 This is a schematic diagram of the CAEL-UNet neural network model structure of the present invention;

[0030] Figure 3 This is a comparison diagram of the differences before and after reconstruction of the random time wind speed component U of the convolutional autoencoder of the present invention.

[0031] Figure 4 This is a comparison diagram of the differences before and after reconstruction of the random time wind speed component V by the convolutional autoencoder of the present invention.

[0032] Figure 5 This is a comparison of the training time required for the CAEL-UNet model and the LSTM-UNet model of this invention under different advance training days.

[0033] Figure 6 This is a comparison of the peak GPU memory usage of the CAEL-UNet model and the LSTM-UNet model during training at different advance training days.

[0034] Figure 7 This is a comparison of the prediction results of the wind speed component U in the CAEL-UNet model and the LSTM-UNet model under different lead times in the present invention;

[0035] Figure 8 This is a comparison of the prediction results of the wind speed component V in the CAEL-UNet model and the LSTM-UNet model under different lead times. Detailed Implementation

[0036] 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.

[0037] like Figures 1-8 As shown, the present invention provides a spatiotemporal synchronous prediction method for wind speed based on CAEL-UNet, and the specific steps are as follows:

[0038] Step 1: Obtain long-term global satellite remote sensing data related to meteorological and environmental elements from publicly available data sources, and extract the easterly wind component data and the northerly wind component data;

[0039] The data source disclosed in this embodiment is the Copernicus Climate Change Service online platform CDS-Beta. CDS-Beta is a climate and environment data platform supported by the European Commission. It can provide users with data in the fields of meteorology, oceanography, land and climate change based on multiple sources such as satellite observations, climate models or ground observation stations. It provides users with convenient data access and download methods, and supports direct download and API batch download.

[0040] The meteorological and environmental element dataset downloaded in this embodiment is wind speed data, which is satellite remote sensing data with long-term time-series characteristics. The wind speed data is divided by latitude and longitude and contains both easterly and northerly wind components. Furthermore, the spatial accuracy of the wind speed data is 0.1°, and the temporal accuracy is 1 hour. The required prediction area data is extracted from the global data range, with the extracted latitude and longitude range being 6.4°.

[0041] Step 2: Preprocess the easterly and northerly wind component data, normalize the preprocessed data, stack the easterly and northerly wind component data from the same time and space to form a spatiotemporal data cube, and then divide the dataset into training and test sets according to time order. The specific process is as follows:

[0042] Step 21: Preprocess the outliers and missing values ​​in the easterly and northerly wind component data using spatial neighborhood filling. Specifically, the "spatial neighboring cells" surrounding the missing values ​​are used as the data source for filling.

[0043]

[0044] In the formula, M represents the grid data of the easterly and northerly wind components, with dimensions of "rows × columns". The result is the filled value. U / D / L / R are the valid values ​​of the upper / lower / left / right neighbors of the outlier and missing value (neighborhoods that are out of bounds or missing are not counted). K is the number of valid neighbors (non-missing values ​​in U / D / L / R), 1≤K≤4.

[0045] Step 22: Normalize the preprocessed data to eliminate dimensional interference, make the data distribution more concentrated, and make the gradient update more stable. The calculation formula is as follows:

[0046]

[0047] In the formula, The minimum value among all samples of a single wind speed data point. Then and on the contrary, For a single sample value to be normalized, These are the normalized sample values. ∈[0,1].

[0048] Step 23: Perform vertical stacking on the normalized data, adding data to the "sample dimension" to increase the total number of samples, while keeping the feature dimension unchanged. Since wind speed data includes both easterly and northerly winds, each occupying a data channel, the wind speed data is packaged into three-dimensional data during the prediction process. Each three-dimensional data includes the easterly and northerly winds at the same time. The data dimension for each hour becomes 64×64×2. Finally, the preprocessed data is divided according to the time scale.

[0049] Step 3: Integrate the convolutional autoencoder into the traditional LSTM-UNet model to construct the CAEL-UNet model. This model adopts a cascaded architecture: the front end is the encoder part of the convolutional autoencoder, which is used for feature compression and dimensionality reduction of the input data; the middle is the LSTM-UNet spatiotemporal prediction module, which is responsible for learning multi-scale spatiotemporal features and temporal dependencies; and the back end is the decoder part of the convolutional autoencoder, which is used for feature reconstruction and spatial resolution restoration. By integrating the convolutional autoencoder into LSTM-UNet, high-dimensional data dimensionality reduction is achieved, which reduces the consumption of computing resources while retaining key spatiotemporal features, thus ensuring the efficiency and accuracy of wind speed prediction as much as possible.

[0050] The convolutional autoencoder employs a symmetrical encoder-decoder architecture for feature compression and reconstruction. The encoder part uses three consecutive convolutional blocks, including 3×3 convolution, LeakyReLU activation, batch normalization, and max pooling operations, to downsample and compress the input 64×64×2 wind speed data into a 32×32×1 bottleneck layer feature representation, achieving an effective compression ratio of 4:1. The convolutional autoencoder introduces Dropout regularization after each convolutional layer, with a Dropout ratio of 0.1-0.2 in the encoder part. The decoder part is symmetrically configured, introducing Dropout regularization and batch normalization operations with a ratio of 0.1-0.2 between each layer, significantly enhancing the model's generalization ability and training stability. In this embodiment, the convolutional autoencoder uses LeakyReLU as the activation function, and the negative slope parameter α is set to 0.1.

[0051] The traditional LSTM-UNet model is an encoder-decoder architecture for spatiotemporal sequence prediction. Its encoder path consists of convolutional and pooling layers, used to extract multi-scale spatial features from each frame of the input sequence layer by layer. At the deepest feature layer of the encoder output, LSTM (Long Short-Term Memory) units are embedded to model the feature sequence temporally and capture its dynamic evolution. The decoder path then gradually restores the spatial resolution through upsampling and convolutional layers, and uses skip connections to fuse high-resolution features from corresponding encoder layers, ultimately achieving a predictive output with both spatiotemporal accuracy.

[0052] like Figure 2 As shown, this embodiment places the encoder part of the convolutional autoencoder at the input end of the traditional LSTM-UNet architecture and the decoder part at the output end of the LSTM-UNet architecture, forming a cascaded processing flow. The encoder part of the convolutional autoencoder is responsible for the initial feature extraction and dimensionality reduction of the original wind speed data. Its output features are then input into the LSTM-UNet for refined spatiotemporal modeling. Finally, the spatiotemporal features output by the LSTM-UNet are used as input, and the decoder of the convolutional autoencoder performs upsampling and feature reconstruction, ultimately generating high-precision prediction results and significantly improving overall computational efficiency. The encoder part of the convolutional autoencoder extracts and downsamples features from each frame of the wind speed data, generating a series of multi-scale spatial feature maps. Subsequently, the obtained multi-scale spatial feature map sequence is input into the LSTM-UNet model. The UNet part is responsible for spatial feature extraction from the input image, extracting features of different scales and levels from the image through convolution and pooling operations. The LSTM part learns the temporal dependencies between different images and outputs a feature sequence that integrates spatiotemporal context information. Finally, the decoder part of the convolutional autoencoder is used to upsample and reconstruct the feature sequence that incorporates spatiotemporal context information to output the final prediction result. Figure 3-4 The figures show the differences before and after reconstruction of the random time wind speed components U and V by the convolutional autoencoder, respectively. The U component represents the east-west wind, and the V component represents the north-south wind. U10 represents the east-west wind at 10m, and V10 represents the north-south wind at 10m.

[0053] In this embodiment, the encoder portion of a convolutional autoencoder compresses the input data of 64×64×2 dimensions into a highly compact 32×32×1 low-dimensional feature representation through convolution and downsampling. This process significantly reduces the data dimensionality while extracting essential features. The decoder is then responsible for reconstructing the original input from this low-dimensional feature. This structure allows it to provide computationally efficient input for downstream tasks when used as a feature extractor, thereby effectively reducing the overall computational resource consumption of the model.

[0054] This invention addresses the challenge of data redundancy in wind speed forecasting by using deep learning to establish an autoencoder model and a prediction model. By mining and constructing more effective feature data from the data, the CAEL-UNet model is built, which significantly reduces the overall computational resource consumption of the model while ensuring the accuracy and effectiveness of wind speed forecasting as much as possible.

[0055] Step 4: Input the training set data into the CAEL-UNet model for learning and training, and use the root mean square error as the loss function for training the model. Evaluate the trained model using the root mean square error, maximum GPU memory usage, and training time; specifically:

[0056] The CAEL-UNet model employs a two-stage training strategy: first, a convolutional autoencoder is pre-trained separately, then fine-tuned jointly with LSTM-UNet. Mean squared error (MSE) is used as the loss function for training the model. The trained model is evaluated using MSE, maximum peak GPU memory usage, and training time as evaluation metrics. The formula for calculating MSE is:

[0057]

[0058] In the formula, RMSE is the root mean square error, and n is the total number of samples. Let be the actual observed wind speed at time point i. This is the predicted wind speed value at the i-th time point output by the model.

[0059] Peak GPU memory usage refers to the highest value of GPU memory usage reached during program execution. This is an important performance monitoring metric, especially in deep learning training and large-scale computational tasks, where model training evaluation is crucial. Figure 5 and Figure 6 As shown.

[0060] Step 5: Input the test set data into the trained CAEL-UNet model for prediction, and evaluate the prediction results using the root mean square error. The prediction results are as follows: Figure 7 and Figure 8 As shown.

[0061] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0062] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A spatiotemporal synchronous prediction method for wind speed based on CAEL-UNet, characterized in that, Includes the following steps: Step 1: Obtain long-term global satellite remote sensing data related to meteorological and environmental elements from publicly available data sources, and extract two sets of wind speed data: easterly wind component data and northerly wind component data. Step 2: Preprocess the easterly wind component data and the northerly wind component data, normalize the preprocessed data, stack the easterly wind component data and the northerly wind component data of the same time and space together to form a spatiotemporal data cube, and then divide the dataset into training set and test set according to time order. Step 3: Integrate the convolutional autoencoder into the traditional LSTM-UNet model to construct the CAEL-UNet model. This model adopts a cascaded architecture: the front end is the encoder part of the convolutional autoencoder, the middle is the LSTM-UNet spatiotemporal prediction module, and the back end is the decoder part of the convolutional autoencoder. Step 4: Input the training set data into the CAEL-UNet model for learning and training, and use the mean squared error as the loss function for training the model; Step 5: Input the test set data into the trained CAEL-UNet model for prediction.

2. The method for spatiotemporal synchronous prediction of wind speed according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step 21: Preprocess the outliers and missing values ​​of the easterly wind component data and the northerly wind component data using the spatial neighborhood filling method. The calculation formula is as follows: In the formula, M represents the grid data of the easterly and northerly wind components, with dimensions of rows × columns. The values ​​are the filled results. U / D / L / R are the valid values ​​of the upper / lower / left / right neighbors of the outlier and missing value locations, and K is the number of valid neighbors, 1≤K≤4; Step 22: Normalize the preprocessed data. The calculation formula is as follows: In the formula, The minimum value among all samples of a single wind speed data point. Then and on the contrary, For a single sample value to be normalized, These are the normalized sample values. ∈[0,1]; Step 23: Stack the normalized data vertically to package the wind speed data into three-dimensional data. Each three-dimensional data includes the easterly wind and northerly wind at the same time. Finally, divide the preprocessed data according to the time scale.

3. The method for spatiotemporal synchronous prediction of wind speed according to claim 1, characterized in that: In step 3, the convolutional autoencoder uses a symmetric encoder-decoder architecture for feature compression and reconstruction. The encoder part uses three consecutive convolutional blocks, including 3×3 convolution, LeakyReLU activation, batch normalization and max pooling operations. The convolutional autoencoder introduces Dropout regularization after each convolutional layer.

4. The method for spatiotemporal synchronous prediction of wind speed according to claim 3, characterized in that: The implementation of the convolutional autoencoder in step 3 is as follows: The encoder part of the convolutional autoencoder extracts and downsamples features from each frame of the wind speed data to generate a series of multi-scale spatial feature maps. Then, the obtained multi-scale spatial feature map sequence is input into the LSTM-UNet model. The UNet part is responsible for extracting spatial features from the input image. Through convolution and pooling operations, features of different scales and levels in the image are extracted. The LSTM part learns the temporal dependencies between different images and outputs a feature sequence that integrates spatiotemporal context information. Finally, the decoder part of the convolutional autoencoder upsamples and reconstructs features from the feature sequence that integrates spatiotemporal context information to output the final prediction result.

5. The method for spatiotemporal synchronous prediction of wind speed according to claim 3, characterized in that: The dropout ratio of the encoder section is 0.1-0.2, and the decoder section is symmetrically configured.

6. The method for spatiotemporal synchronous prediction of wind speed according to claim 1, characterized in that: The CAEL-UNet model employs a two-stage training strategy: first, pre-train the convolutional autoencoder separately, and then fine-tune it jointly with LSTM-UNet.

Images