A short-term wind speed prediction method, system, storage medium and device

Abstract

The application discloses a short-term wind speed prediction method and system, a storage medium and equipment. By obtaining the original wind speed sequence after denoising, noise interference is suppressed, the original data is screened before feature extraction, the optimal historical wind speed feature is obtained, the input redundancy and calculation burden are reduced, the multi-scale feature in the optimal historical wind speed feature is extracted, and the multi-scale features are fused, so that the complex fluctuation modes such as high-frequency turbulence, medium-frequency gust and low-frequency trend in the wind speed sequence are fully captured, and the representation ability of the model to the dynamic characteristics of the wind speed is enhanced. In the prediction stage, the fused features are converted to the frequency domain for blocking and coding, and the time domain label and the prediction value are converted to the frequency domain. Through the conversion between the frequency domain and the time domain, the self-correlation problem of the label sequence is solved, and the prediction accuracy of the model is improved.

Description

Technical Field

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

[0002] Under the global trend of energy transition, wind power, as a core clean and renewable energy source, has seen explosive growth in installed capacity. However, the intermittency, volatility, and randomness of wind speed remain key bottlenecks restricting large-scale grid connection of wind power and stable grid operation. Real-time mismatch between wind power output and electricity demand not only causes wind power curtailment during low-demand periods but also leads to power shortages during peak-demand periods, seriously affecting the economic efficiency and reliability of the power system. Therefore, high-precision and high-efficiency short-term wind speed forecasting technology has become a core requirement for optimizing wind farm operation and supporting power dispatching decisions.

[0003] However, current mainstream short-term wind speed forecasting methods all have significant limitations: physical models, such as numerical weather prediction (NWP) and computational fluid dynamics (CFD), while capable of simulating and capturing long-term meteorological trends through physical equations, suffer from extremely high computational costs due to their reliance on complex parameter settings and massive data inputs. Furthermore, they lack sufficient accuracy in characterizing local micro-meteorological conditions, failing to meet the demands for minute-level real-time forecasts. Statistical models, such as autoregressive integral moving average (ARIMA) and exponential smoothing (ES), while simple in principle, computationally inexpensive, and easy to implement in engineering, are essentially linear models. They cannot capture the complex nonlinear dynamic characteristics of wind speed sequences and are highly sensitive to data noise, resulting in a significant drop in prediction accuracy in noisy real-world data environments. While artificial intelligence models have become mainstream due to their powerful nonlinear fitting capabilities, shallow models, such as backpropagation neural networks (BPNN) and support vector machines (SVM), are limited by network depth and feature extraction capabilities, making it difficult to handle long-term dependence and multi-scale variations in wind speed. While deep models such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer models have made breakthroughs in feature extraction and global modeling, they still face many challenges. First, noise in the raw wind speed data can easily lead to the model learning false features. Existing denoising methods based on Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD) not only introduce additional computational costs but also easily misclassify noise as valid signals. Second, the high complexity of the Transformer model's self-attention mechanism, coupled with the computational cost of signal decomposition, results in excessively long model training times, making it difficult to meet the engineering requirements of periodically updating models in wind farms. Third, the Transformer's direct prediction paradigm ignores the inherent autocorrelation of wind speed label sequences, and traditional loss functions do not adapt to this, leading to training bias. Fourth, existing models mostly model from the time domain, failing to directly capture the multi-frequency component features of wind speed and thus unable to comprehensively depict the dynamic changes in wind speed.

[0004] In summary, existing short-term wind speed prediction methods all suffer from problems such as an imbalance between accuracy and efficiency, weak noise resistance, and insufficient characterization of wind speed sequence characteristics. They cannot simultaneously meet the multiple requirements of actual wind farms for prediction accuracy, training efficiency, and engineering practicality. Developing a short-term wind speed prediction technology with low training cost, noise resistance, and high accuracy has become an urgent task for the development of the wind power industry. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of prediction bias caused by autocorrelation of label sequences and insufficient learning of frequency features in the prior art, which leads to low accuracy of prediction results, and to provide a short-term wind speed prediction method, system, storage medium and device.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A short-term wind speed forecasting method includes the following steps: Obtain the original wind speed sequence after denoising, and select the best historical wind speed features from the original wind speed sequence after denoising. Multi-scale features are extracted from the best historical wind speed features, and the multi-scale features are fused to obtain the fused features. The fused features are converted to the frequency domain to obtain frequency domain coefficients. The frequency domain coefficients are divided into multiple frequency subsequences, and the frequency subsequences are encoded with frequency features. The encoded frequency features are converted back to the time domain to obtain the wind speed prediction result.

[0007] A further improvement of the present invention is that: The process of selecting the optimal historical wind speed features from the denoised original wind speed sequence includes the following steps: the Spearman correlation coefficient between historical data with different lag steps and the target wind speed in the denoised wind speed sequence. Calculate the Spearman correlation coefficient between historical data with different lag steps and the target wind speed in the denoised wind speed sequence; By setting a correlation coefficient threshold and using the Spearman correlation coefficient between historical data with different lag steps and the target wind speed, the corresponding historical wind speed data is selected to obtain the optimal historical wind speed characteristics.

[0008] The process involves extracting multi-scale features from the optimal historical wind speed characteristics and fusing these multi-scale features to obtain fused features, including: The optimal historical wind speed feature is convolved using multiple convolution kernels of different scales to obtain convolution features of different scales. A squeeze-excitation mechanism is introduced, and the weights of each convolutional feature are dynamically calculated based on the squeeze-excitation mechanism; Based on the weight of each convolutional feature, multiple convolutional features of different scales are weighted and fused to obtain the fused features.

[0009] The fused features are transformed to the frequency domain to obtain frequency domain coefficients, which are then divided into multiple frequency sub-sequences, including: The fused features are transformed to the frequency domain using discrete Fourier transform, and frequency coefficients are extracted. The frequency coefficients are divided into N non-overlapping local frequency subsequences; The subsequence is normalized to obtain the normalized subsequence.

[0010] The frequency feature encoding of the frequency subsequence includes: Position codes are added to each frequency subsequence to supplement timing information; The subsequences with added positional encoding are input into the multi-head self-attention module, and the multi-head attention output and linear transformation are concatenated to obtain the final attention output.

[0011] The step of converting the encoded frequency features back to the time domain to obtain the wind speed prediction result includes: The encoded frequency domain subsequence The concatenated frequency coefficients are transformed back to the time domain using an inverse discrete Fourier transform to obtain the prediction result:

[0012] in, For linear projection layers; T is the prediction step size; This represents the encoded frequency domain subsequence; n represents the time domain sequence index; L represents the sequence length; Indicates the frequency domain sequence index; A short-term wind speed forecasting system includes: The feature filtering module is used to obtain the denoised original wind speed sequence and filter out the best historical wind speed features from the denoised original wind speed sequence. The feature fusion module is used to extract multi-scale features from the best historical wind speed features and fuse the multi-scale features to obtain the fused features. The prediction module is used to convert the fused features to the frequency domain to obtain frequency domain coefficients, divide the frequency domain coefficients into multiple frequency subsequences, encode the frequency features of the frequency subsequences, and convert the encoded frequency features back to the time domain to obtain the wind speed prediction result.

[0013] A computer program product includes a computer program that, when executed by a processor, implements the method described in any one of the present invention.

[0014] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described in this invention.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods described in this invention.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a short-term wind speed prediction method. By acquiring the denoised original wind speed sequence, noise interference is suppressed. Before feature extraction, the original data is screened to obtain the optimal historical wind speed features, reducing input redundancy and computational burden. Multi-scale features are extracted from the optimal historical wind speed features and fused to fully capture complex fluctuation patterns such as high-frequency turbulence, mid-frequency gusts, and low-frequency trends in the wind speed sequence, enhancing the model's ability to represent the dynamic characteristics of wind speed. In the prediction stage, the fused features are converted to the frequency domain for block segmentation and encoding, and the time domain labels and predicted values ​​are converted to the frequency domain. Through the conversion between the frequency domain and the time domain, the autocorrelation problem of the label sequence is solved, improving the prediction accuracy of the model. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart disclosed in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0021] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0022] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0024] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1This invention discloses a short-term wind speed prediction method. Existing short-term wind speed prediction methods suffer from technical pain points such as high data noise interference, high training computation costs, prediction bias due to label sequence autocorrelation, and insufficient frequency feature learning. This invention addresses these issues by employing StOMP denoising, dynamic convolution (DC) multi-scale feature fusion, frequency domain debiasing Transformer, and frequency domain decorrelation (FDD) loss function. It constructs a fully integrated short-term wind speed prediction technology system encompassing "data preprocessing - multi-scale feature fusion - frequency domain modeling - loss function optimization," reducing training time while maintaining prediction accuracy to meet the application requirements of wind farms. Each module is deeply coupled through mathematical modeling and algorithm design to achieve synergistic optimization of "denoising - efficiency improvement - bias reduction - high accuracy," specifically including the following steps: Example 1 This invention discloses a short-term wind speed forecasting method, comprising the following steps: Step 1: Obtain the denoised original wind speed sequence, and select the best historical wind speed features from the denoised original wind speed sequence; Step 2: Extract multi-scale features from the best historical wind speed data and fuse the multi-scale features to obtain the fused features; Step 3: Convert the fused features to the frequency domain to obtain frequency domain coefficients. Divide the frequency domain coefficients into multiple frequency subsequences, encode the frequency features of the frequency subsequences, and convert the encoded frequency features back to the time domain to obtain the wind speed prediction result.

[0026] Example 2 Step 1: Data preprocessing, StOMP denoising and Spearman feature selection This module addresses the issues of noise interference and redundant input features in the original wind speed data by using sparse representation for noise reduction and relevant feature selection, thus providing a high-quality data foundation for subsequent model training.

[0027] Step 1.1, StOMP noise reduction Based on sparse representation theory, assuming the original wind speed sequence consists of effective signals and non-sparse noise, signal and noise separation is achieved through atomic filtering using an overcomplete dictionary. The specific formula is as follows: Signal modeling: Let the 1D original wind speed sequence be... Where m is the sequence length, its mathematical expression is:

[0028] in, For an overcomplete dictionary, it is composed of Gaussian wavelet atoms composition, , Let be a sparse coefficient vector, satisfying , This refers to non-sparse noise, such as sensor errors and electromagnetic interference.

[0029] Iterative Atom Selection: This involves multiple rounds of iterative selection to find dictionary atoms that match the valid signal. The specific process is as follows: Initialization: Residual The initial residual is equal to the original sequence, and the sparsity coefficient is... Supports collection , where there are no selected atoms initially, and the number of iterations. ; Atomic Sieve (Sift): Calculates the inner product of the residual and all dictionary atoms, and filters out those whose absolute value of the inner product exceeds a threshold. Atomic indexes, update the support set:

[0030] in, This represents the inner product operation. The index of the atom selected in the k-th round; Sparse coefficient update (Least Square Estimation): based on support set Corresponding sub-dictionary The optimal sparsity coefficients are obtained by least squares estimation:

[0031] in, Describing the L2 norm, This is the transpose matrix of the sub-dictionary; Residual Update: Calculate the new residuals to evaluate the effect of the iteration.

[0032] Iteration terminates: when or number of iterations The maximum number of iterations is preset, and the embodiment of the present invention takes... When the iteration stops, zero-padding is used to fill the gaps. Expanded to full sparse coefficients The final output is a denoised sequence:

[0033] Key parameter settings: The overcomplete dictionary D uses a Gaussian wavelet dictionary (atom length matches the wind speed sequence sampling interval, set to 16-32), and the threshold... The invention adaptively adjusts based on noise intensity; specifically, it takes... This ensures that effective fluctuation characteristics are preserved while suppressing noise.

[0034] Step 1.2, Spearman Relevance Feature Selection Wind speed sequences exhibit time-dependent relationships, specifically a correlation between historical and future data. The Spearman coefficient is used to quantify this correlation under different lag steps, allowing for the selection of optimal input features and reducing redundant computation. The formula is as follows: Spearman coefficient definition: Let the historical wind speed sequence be... L is the maximum hysteresis step size, which is taken in this invention. Target wind speed is Calculate the Spearman correlation coefficient between the two:

[0035] Where m is the sample size. The difference in rank, This indicates the rank of the variable in the sequence. , The larger the absolute value, the stronger the correlation.

[0036] Feature selection rules: Retain Historical data corresponding to the lag step size can be used as model input. For example, 24 steps of historical data can be selected from the Samoa (SMO) dataset, 23 steps from Barrow (BRW) dataset, and 12 steps from Mauna Loa (MLO) dataset. The input feature dimension is reduced by 30%-60%, which significantly improves the efficiency of subsequent model training.

[0037] Step 2, Multi-scale feature fusion, Dynamic Convolution (DC) design To address the multi-scale fluctuation characteristics of wind speed sequences, such as high-frequency turbulence, mid-frequency gusts, and low-frequency trends, traditional fixed convolutional kernels cannot adaptively capture features at different scales. This invention designs a dynamic convolution module that adjusts the convolutional kernel combination by adjusting the input-dependent attention weights, achieving efficient fusion of multi-scale features. The specific formula is as follows: Step 2.1, Modeling Dynamic Convolutional Kernels and Attention Weights Kernel settings: Introduce K 1D convolution kernels of different scales. Specifically, the kernel sizes are 2, 3, 4, and 5, covering the main fluctuation scales of wind speed, and each convolution kernel corresponds to a weight. Where A is the frequency domain coefficient vector, satisfying: Weight This represents the contribution of the k-th convolutional kernel to the current input, dynamically calculated by the squeeze-excitation (SE) mechanism.

[0038] SE mechanism calculates attention weights: Global Average Pooling (GAP): Extracts global statistical features of frequency domain coefficients A, compressing spatial dimensions.

[0039] Where L is the length of A Let A be the l-th element. It contains global information about both the real and imaginary parts; Dimensionality transformation: By using fully connected layers to reduce and increase dimensionality, the dependencies between feature channels are learned.

[0040]

[0041] in, r is the dimensionality reduction coefficient, which is taken in this invention. , For dimensionality reduction bias; For dimensional bias; It is the ReLU activation function; Softmax normalization: Outputs the final attention weights. in, Let v be the k-th element, and ensure that the sum of the weights is 1.

[0042] Step 2.2, Multi-scale feature fusion A serial skip-layer connection structure is adopted to fuse features from different convolutional layers, avoiding information loss in traditional serial structures. The formula is as follows:

[0043] in, This is the output of the l-th dynamic convolution layer. This represents a 1D convolution operation. It is the ReLU activation function. These represent the kernel and bias of the l-th layer, respectively. For the first Layer output (skip-layer connection) retains both shallow local features and deep global features through this structure, thereby improving feature representation capabilities.

[0044] Step 3, Frequency Domain Modeling Module: Frequency Domain De-biasing Transformer Traditional Transformers model wind speed in the time domain, failing to directly capture its multi-frequency components, such as intraday fluctuations and seasonal trends. Furthermore, their self-attention mechanism overemphasizes high-amplitude frequency components, leading to frequency bias. This invention achieves unbiased learning of frequency features through a process of "time-frequency domain transformation - frequency block normalization - Transformer encoding - frequency-time domain inverse transformation," as detailed in the following formula: Step 3.1, Time-domain to frequency-domain conversion The denoised wind speed sequence Where L is the sequence length, the sequence is transformed to the frequency domain using Discrete Fourier Transform (DFT) to extract the frequency coefficients.

[0045] in, Here is the frequency domain coefficient vector; j is the imaginary unit, which can be decomposed into real part moments. With the imaginary part matrix The coefficient represents the k-th frequency component, reflecting the amplitude and phase information of that frequency.

[0046] Step 3.2, Frequency Blocking and Normalization Frequency partitioning: Dividing A into N non-overlapping local frequency subsequences solves the efficiency problem of Transformer in processing long sequences.

[0047] in, The length of the subsequence is taken in this invention. N is the number of subsequences. By dividing the long sequence into blocks, the complexity of the self-attention mechanism is reduced.

[0048] Min-Max normalization: eliminates amplitude differences between subsequences of different frequencies, preventing the model from overemphasizing high-amplitude components.

[0049] in, Let i be the normalized i-th subsequence, and its value range is 1. These represent the minimum and maximum values ​​of the subsequence, respectively.

[0050] Step 3.3, Position Encoding and Transformer Encoder Location encoding: Transformer has no temporal dependency, and supplements the temporal information of the wind speed sequence with sine and cosine encoding:

[0051]

[0052] Where pos is the sequence position index. d is the position encoding dimension, which is taken in this invention. ), where i is the dimension index , For the position encoding matrix, and the normalized subsequence The spliced ​​data is used as the encoder input.

[0053] Multi-head self-attention mechanism: Learns independently for each normalized subsequence to avoid cross-frequency interference, as shown in the following formula: Query / Key / Value projection: Projecting the input Generate Query(Q), Key(K), and Value(V) through linear transformation:

[0054] in, For learnable parameter matrix, m is the number of attention heads, which is taken in this invention. ; Attention score calculation:

[0055] in, This is a scaling factor to prevent the softmax gradient from vanishing due to excessively large scores. Multi-head splicing: The outputs of m attention heads are spliced ​​together and linearly projected to obtain the final attention output.

[0056] in, For the projection parameter matrix, This is the attention output for the i-th subsequence.

[0057] Encoder output: Each Transformer encoder layer contains a "multi-head attention + feedforward neural network (FFN)", the FFN formula is as follows:

[0058] in, For the FFN parameter matrix, For bias, This is the ReLU activation function.

[0059] After two encoder layers, the encoded results of all subsequences are obtained:

[0060] Step 3.4, frequency domain to time domain inverse conversion The encoded frequency domain subsequence The concatenated frequency coefficients are then transformed back to the time domain using the inverse discrete Fourier transform (IDFT) to obtain preliminary prediction results.

[0061] in, For linear projection layers, Dimension mapped to L, T is the prediction step size, which is taken in this invention. That is, single-step prediction.

[0062] Step 4, Loss Function Optimization Module: Frequency Domain Decorrelation (FDD) Loss Traditional MSE loss ignores the autocorrelation of wind speed label sequences. This can lead to model training bias due to the orthogonality of frequency domain components. This invention utilizes the orthogonality of frequency domain components to design an FDD loss function to reduce autocorrelation interference, as shown in the following formula: Step 4.1, Time-domain to frequency-domain conversion Using Fast Fourier Transform (FFT), the complexity is... ) Time-domain label sequence With predicted sequence Convert to frequency domain:

[0063]

[0064] in, These are the k-th frequency domain coefficients of the label and the predicted sequence, respectively. Since the frequency domain components are orthogonal... This can eliminate time-domain autocorrelation.

[0065] Step 4.2, Define the FDD loss function Using the L1 norm of the frequency domain modulus as the loss function avoids amplifying low-frequency errors by using the squared term of the MSE:

[0066] in, Modulus of a complex number The L1 norm can alleviate the problem of low-frequency error amplification in MSE and improve training stability.

[0067] Step 5: The entire process of model training and prediction Combining the above modules, construct The model, training, and prediction process is as follows: Data partitioning: The wind speed data is divided into 80% (training set), 5% (validation set), and 15% (test set). The training set is used for model parameter learning, the validation set is used for hyperparameter optimization (such as learning rate and batch size), and the test set is used for performance evaluation. Hyperparameter settings: Optimizer selected: Adam; Learning rate: ,batch Number of training rounds , , .

[0068] Training process: Gradient descent minimization is employed. Gradient explosion is avoided by gradient clipping, and hyperparameters are adjusted using a validation set after each iteration. Predicted Output: Input the test set into the trained model, and it outputs short-term wind speed prediction results. Performance is evaluated using RMSE, MMAPE, and spectral error (SE), and the formulas are as follows:

[0069]

[0070]

[0071]

[0072] Where N is the number of test samples, These are the actual value and the predicted value, respectively. The mean of the true values ​​is used to evaluate the accuracy of frequency domain prediction.

[0073] Example 3 This invention also discloses a short-term wind speed forecasting system, comprising: The feature filtering module is used to obtain the denoised original wind speed sequence and filter out the best historical wind speed features from the denoised original wind speed sequence. The feature fusion module is used to extract multi-scale features from the best historical wind speed features and fuse the multi-scale features to obtain the fused features. The prediction module is used to convert the fused features to the frequency domain to obtain frequency domain coefficients, divide the frequency domain coefficients into multiple frequency subsequences, encode the frequency features of the frequency subsequences, and convert the encoded frequency features back to the time domain to obtain the wind speed prediction result.

[0074] This invention proposes a short-term wind speed prediction method, constructing a complete process system of "data preprocessing - multi-scale feature fusion - frequency domain modeling - loss function optimization". Based on sparse representation theory, the StOMP algorithm is used to denoise the original wind speed sequence. Historical data selected by Spearman correlation coefficient is used as input features to reduce redundancy. An input-dependent dynamic convolution module is designed, and attention weights based on the SE mechanism are used to adaptively adjust the combination of multi-scale convolution kernels to avoid information loss. The denoised time-domain sequence is transformed to the frequency domain. After learning multi-frequency features through a frequency-locally independent Transformer encoder combined with a multi-head attention mechanism, the sequence is transformed back to the time domain by IDFT to obtain the preliminary prediction result. To solve the problem of autocorrelation in the label sequence, FFT is used to transform the time-domain labels and predicted values ​​to the frequency domain. An FDD loss function in the form of frequency domain modulo L1 norm is designed to reduce autocorrelation interference and improve training stability. The model training results are finally evaluated using metrics such as R² and RMSE. This scheme achieves low training cost, noise resistance, and high accuracy in short-term wind speed prediction by replacing traditional decomposition with StOMP denoising, dynamically convolutional fusion of multi-scale features, frequency domain debiased Transformer modeling, and FDD loss function optimization.

[0075] A schematic diagram of a terminal device according to an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0076] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0077] The terminal device can be a desktop computer, laptop computer, cloud server, or other device with strong computing power. The terminal device may include, but is not limited to, a processor and memory.

[0078] The optimal choice for the processor is a multi-core high-speed central processing unit (CPU).

[0079] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0080] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0081] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A short-term wind speed prediction method, characterized in that, Includes the following steps: Obtain the original wind speed sequence after denoising, and select the best historical wind speed features from the original wind speed sequence after denoising. Multi-scale features are extracted from the best historical wind speed features, and the multi-scale features are fused to obtain the fused features. The fused features are converted to the frequency domain to obtain frequency domain coefficients. The frequency domain coefficients are divided into multiple frequency subsequences, and the frequency subsequences are encoded with frequency features. The encoded frequency features are converted back to the time domain to obtain the wind speed prediction result.

2. The short-term wind speed prediction method according to claim 1, characterized in that, The process of selecting the optimal historical wind speed features from the denoised original wind speed sequence includes the following steps: the Spearman correlation coefficient between historical data with different lag steps and the target wind speed in the denoised wind speed sequence. Calculate the Spearman correlation coefficient between historical data with different lag steps and the target wind speed in the denoised wind speed sequence; By setting a correlation coefficient threshold and using the Spearman correlation coefficient between historical data with different lag steps and the target wind speed, the corresponding historical wind speed data is selected to obtain the optimal historical wind speed characteristics.

3. The short-term wind speed prediction method according to claim 1, characterized in that, The process involves extracting multi-scale features from the optimal historical wind speed characteristics and fusing these multi-scale features to obtain fused features, including: The optimal historical wind speed feature is convolved using multiple convolution kernels of different scales to obtain convolution features of different scales. A squeeze-excitation mechanism is introduced, and the weights of each convolutional feature are dynamically calculated based on the squeeze-excitation mechanism; Based on the weight of each convolutional feature, multiple convolutional features of different scales are weighted and fused to obtain the fused features.

4. The short-term wind speed prediction method according to claim 1, characterized in that, The fused features are transformed to the frequency domain to obtain frequency domain coefficients, which are then divided into multiple frequency sub-sequences, including: The fused features are transformed to the frequency domain using discrete Fourier transform, and frequency coefficients are extracted. The frequency coefficients are divided into N non-overlapping local frequency subsequences; The subsequence is normalized to obtain the normalized subsequence.

5. The short-term wind speed prediction method according to claim 4, characterized in that, The frequency feature encoding of the frequency subsequence includes: Position codes are added to each frequency subsequence to supplement timing information; The subsequences with added positional encoding are input into the multi-head self-attention module, and the multi-head attention output and linear transformation are concatenated to obtain the final attention output.

6. The short-term wind speed prediction method according to claim 4, characterized in that, The step of converting the encoded frequency features back to the time domain to obtain the wind speed prediction result includes: The encoded frequency domain subsequence The concatenated frequency coefficients are transformed back to the time domain using an inverse discrete Fourier transform to obtain the prediction result: in, For linear projection layers; T is the prediction step size; This represents the encoded frequency domain subsequence; n represents the time domain sequence index; L represents the sequence length; This represents the frequency domain sequence index.

7. A short-term wind speed forecasting system, characterized in that, include: The feature filtering module is used to obtain the denoised original wind speed sequence and filter out the best historical wind speed features from the denoised original wind speed sequence. The feature fusion module is used to extract multi-scale features from the best historical wind speed features and fuse the multi-scale features to obtain the fused features. The prediction module is used to convert the fused features to the frequency domain to obtain frequency domain coefficients, divide the frequency domain coefficients into multiple frequency subsequences, encode the frequency features of the frequency subsequences, and convert the encoded frequency features back to the time domain to obtain the wind speed prediction result.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images