Wind speed prediction method based on conditional residual gating circulation unit model
By employing a wind speed prediction method based on a conditional residual gated cyclic unit model, utilizing VMD decomposition and multi-head self-attention mechanism, and combining NWP wind speed information, the problem of insufficient wind speed prediction accuracy in existing technologies is solved, achieving higher accuracy in wind speed and wind power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-15
AI Technical Summary
Existing wind speed prediction methods that rely on NWP data lack in-depth analysis of the multi-scale dynamic components of measured wind speed sequences, making it difficult to simultaneously consider both trends and fluctuations in prediction under the constraints of low-resolution exogenous information.
A wind speed prediction method based on the Conditional Residual Gated Recurrent Unit (Con-ResGRU) model is adopted. The measured wind speed sequence is decomposed by the parameter-optimized VMD module. Combined with the multi-head self-attention mechanism and residual connection, the future wind speed is predicted step by step by combining the historical measured wind speed and the NWP wind speed. A hybrid loss function is used to guide the training.
It significantly improves the accuracy and stability of wind speed forecasting, and can better capture the multi-scale dynamic characteristics and trends of wind speed, providing more reliable wind power forecasting input.
Smart Images
Figure CN122046964A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind speed prediction technology, and in particular to a wind speed prediction method based on a conditional residual gated cyclic unit model. Background Technology
[0002] Wind energy is an important component of renewable energy. Influenced by factors such as pressure, topography, and temperature, wind is an extremely unstable natural resource, characterized by randomness, intermittency, and uncertainty, making it one of the most difficult meteorological elements to predict. Inaccurate wind speed predictions further affect the accuracy of wind power forecasting, posing a challenge to grid dispatch. Accurate wind speed forecasts not only provide a scientific basis for wind farm operation and management, improving the stability and reliability of wind power grid connection, but also provide important references for electricity market transactions, energy dispatch, and wind energy resource assessment. However, the inherent characteristics of wind make processing wind speed sequences and analyzing wind speed characteristics a significant challenge in improving forecast accuracy. Therefore, how to effectively process wind speed data, extract wind speed characteristics, and improve forecast accuracy has become one of the current research hotspots in the wind energy field. Existing wind power forecasting methods can be divided into ultra-short-term, short-term, and medium-to-long-term forecasts according to the forecast time scale. Ultra-short-term and short-term forecasts mainly provide information for real-time grid dispatch and provide a basis for power plants to specify daily power generation plans; while medium-to-long-term forecasts support tasks such as wind farm site selection and turbine maintenance. Because wind speed has strong random fluctuations, especially in the short term, it has a significant impact on the power system. Therefore, current research in this field mainly focuses on ultra-short-term and short-term forecasting.
[0003] Currently, artificial intelligence technology has become a research hotspot in the field of wind speed and wind power prediction. Commonly used models include support vector machines, autoencoders, convolutional neural networks, and recurrent neural networks. However, due to the strong nonlinearity and non-stationarity of wind speed itself, these single models struggle to capture all the characteristics of the original wind speed. Therefore, many scholars have made numerous attempts to improve the performance of wind speed prediction models, manifested in the transition from mathematical statistical models to machine learning and deep learning, and the development from single models to coupled models. Among these, decomposition-coupled models are the most widely studied. These models decompose the original sequence into a series of stationary subsequences before model training. Common decomposition methods include Fourier transform, wavelet transform, empirical mode decomposition, and variational mode decomposition. These subsequences are then input into the coupled prediction model, effectively improving the accuracy of wind speed prediction.
[0004] Wind power forecasting typically relies on NWP (Numerical Weather Prediction) data. However, due to limitations in the spatiotemporal resolution and the highly simplistic nature of NWP data, the actual wind speed for individual wind turbines within a wind farm is significantly affected by topography, wake effects, and other factors, resulting in a substantial deviation from the wind speed data predicted by numerical weather prediction. This reduces the accuracy of direct power prediction based on NWP wind speed data. Existing research on improving the accuracy of NWP wind speed data largely focuses on post-processing of the data itself, lacking in-depth analysis of the multi-scale dynamic components of measured wind speed sequences. This makes it difficult to simultaneously consider both trends and fluctuations in forecasting under low-resolution exogenous information constraints. Therefore, it is necessary to design a wind speed forecasting scheme that can simultaneously utilize multi-scale information from measured wind speeds and future information from NWP wind speed data, while possessing higher prediction accuracy and stability. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a wind speed prediction method based on a conditional residual gated cyclic unit model. This method solves the technical problem that existing wind speed prediction methods that rely on NWP data lack in-depth analysis of the multi-scale dynamic components of measured wind speed sequences and struggle to simultaneously assess trends and fluctuations under low-resolution exogenous information constraints.
[0006] The technical solution adopted in this invention is as follows: This invention provides a wind speed prediction method based on a conditional residual gated cyclic unit model, comprising: Obtain the historical measured wind speed sequence of the target wind turbine and the corresponding NWP wind speed sequence for the time period, and at the same time obtain the NWP future wind speed sequence for the prediction period; Data processing is performed on each wind speed sequence. The historical measured wind speed sequence is decomposed using the parameter-optimized VMD module to obtain several intrinsic mode function components. These components, together with the remaining wind speed sequences, constitute a dataset, which is then divided into a training set, a test set, and a validation set. The Con-ResGRU conditional residual gated recurrent neural network model is trained using the training and validation sets to obtain the prediction model; The Con-ResGRU comprises an encoder, a multi-head self-attention mechanism, and a decoder arranged sequentially. The encoder employs a ResGRU model, which uses a multi-channel feature vector formed by concatenating the intrinsic mode function components in the time dimension, along with the NWP wind speed sequence input, to initially extract the input temporal features and capture time-dependent information. The multi-head self-attention mechanism is used to globally model the encoder output, thereby enhancing the expression of long-range dependencies between different time steps of the historical measured wind speed sequence. The decoder, based on the ResGRU model, uses the output of the multi-head self-attention mechanism as the state input and the future NWP wind speed as the conditional input to incorporate into each prediction step, and obtains the predicted wind speed for the future time step through output layer mapping. The test set is input into the prediction model to obtain the wind speed prediction results.
[0007] The preferred technical solution is: The VMD module with optimized parameters decomposes the historical measured wind speed sequence, including: Using genetic algorithms to analyze the key parameter of VMD: number of modes K With bandwidth factor α Joint optimization is performed, and the objective function is: , In the formula, These are the components of each intrinsic mode function; The center frequency of each intrinsic mode function component; It is a pulse function; Modal number; For the first The modes obtained from the decomposition t Time is the independent variable; The exponent is a complex number, representing frequency shift and demodulation; The original signal; The energy entropy of the residual signal obtained by decomposing the historical measured wind speed sequence using VMD is used as the fitness function to measure the randomness and information concentration of the decomposed signal. Inconsistent interpretations are penalized through parameter boundary and integer constraint mechanisms. Finally, the optimal values of key parameters and the K intrinsic mode function components and decomposition residuals obtained after decomposition are output.
[0008] When the parameter-optimized VMD module decomposes the historical measured wind speed sequence, it constructs rolling standard deviation volatility characteristics with different window lengths for the historical measured wind speed sequence, making it easier for the prediction model to identify sudden wind speed changes and make reasonable predictions.
[0009] The encoder uses a ResGRU model consisting of two vertically stacked GRUs, and the gating update process includes: , in, To update the door, To reset the door, , For the prediction period t Hidden states and candidate hidden states at each time step Representing the t- Decoding status at one time step, For the Sigmoid function, For Hadamard products, , , This represents the weight matrix and bias.
[0010] The decoder outputs the first... t Hidden state at each time step As an initial state, at each time step, the decoder's input contains not only the hidden state from the previous time step. It also explicitly introduces the future wind speed of NWP at the corresponding time as a condition input, and then outputs the prediction results for the prediction period step by step to realize the dynamic correction of the future wind speed of NWP.
[0011] The multi-head self-attention mechanism incorporates residual connections and a feedforward network; the feedforward network is a two-layer fully connected network used to perform nonlinear feature reconstruction on the attention output.
[0012] The multi-head self-attention mechanism performs global modeling of the encoder output, including:
[0013] in, The hidden state output by the encoder. for With multi-head self-attention output After summing the residuals, perform layer normalization. The obtained intermediate feature sequence, Feedforward network The nonlinear transformation output; The attention-enhanced feature sequence serves as the input context for the encoder and decoder.
[0014] The Con-ResGRU is configured with a global residual connection, and the decoder output outputs the predicted wind speed for the future time step through the global residual connection.
[0015] The prediction model uses a hybrid loss function that integrates mean squared error and volatility loss term to guide parameter updates; The hybrid loss function ,in: , , , In the formula, α and β are weighting coefficients. Mean square error, and The prediction model is respectively in the th t The, the t- Predicted wind speed output at one time step. and These are the corresponding actual wind speeds; H This is the total prediction step size. This is the volatility loss term for the first-order difference. and The first t The first-order difference between the actual wind speed and the predicted wind speed at each time step is used to balance the importance of numerical error and volatility loss constraints.
[0016] The prediction results output by the prediction model are expressed as follows: , As a baseline forecast term This is the residual correction amount for the model output.
[0017] The technical solution of the present invention can achieve at least some of the following beneficial effects: Compared to directly using NWP wind speed for prediction, this invention can more fully exploit the multi-scale dynamic components of measured wind speed, improving the accuracy of wind speed prediction under low spatiotemporal resolution NWP conditions, and providing input for high-precision wind power prediction. Specifically, this invention has the following advantages: This invention uses VMD decomposition to help extract different frequency-level features of historical measured wind speeds, remove non-stationarity and multi-scale aliasing in historical sequences, enable the model to better capture trends and fluctuations, and significantly improve the decomposition quality through adaptive parameter optimization.
[0018] This invention utilizes the combination of historical measured wind speeds and corresponding historical wind speeds of the NWP (Near-Wave Power) to help the model learn the systematic deviation between the NWP and the actual wind speed. During the decoding stage, when predicting future sequences step by step, the future wind speed of the NWP corresponding to the prediction period is explicitly used as a conditional input. This ensures that each output is constrained by the meteorological priors at the corresponding time, making the prediction results closer to the actual wind speed under the combined effect of "historical multi-scale patterns" and "future forecast information". This greatly improves the prediction accuracy and provides a more reliable wind speed input basis for subsequent high-precision power prediction.
[0019] This invention introduces global residual connections in Con-ResGRU to provide a stable baseline for prediction and mitigate gradient vanishing, thereby improving the training stability and convergence speed of deep networks.
[0020] This invention constructs multi-scale volatility features at the input end and employs a hybrid loss function at the training end, thereby enhancing the fitting ability and prediction smoothness for regions with sudden changes in wind speed.
[0021] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description
[0022] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the Con-ResGRU conditional residual gated recurrent neural network model in an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of the ResGRU model in an embodiment of the present invention.
[0025] Figure 4 This is a schematic diagram of the basic framework of the multi-head self-attention mechanism in an embodiment of the present invention.
[0026] Figure 5 The prediction results of the prediction model in this embodiment of the invention are compared with those of other schemes. Detailed Implementation
[0027] The specific embodiments of the present invention are described below with reference to the accompanying drawings.
[0028] See Figure 1 This embodiment provides a wind speed prediction method based on a conditional residual gated cyclic unit model, using SCADA data (real-time production data collected and monitored by the SCADA system) and NWP data from a wind turbine in Jiangsu Province. The method includes the following steps: S1. Obtain the historical measured wind speed sequence of the target wind turbine and the corresponding NWP wind speed sequence from the SCADA data and NWP data, and at the same time obtain the future NWP wind speed sequence for the predicted period.
[0029] S2. Process the data of each wind speed sequence. Use the parameter-optimized VMD (Variational Mode Decomposition) module to decompose the historical measured wind speed sequence to obtain several intrinsic mode function components. These components, together with the remaining wind speed sequences, constitute a dataset. Then, divide the dataset into training set, test set, and validation set.
[0030] Specifically, the data processing of each sequence includes: First, the raw data is cleaned: this includes aligning the time granularity, handling zero values, missing values, and outliers to eliminate the impact of bad data and ensure the accuracy of the data input into the model. Then, the isolated forest algorithm is used to identify outliers, and the median under the same power is used as a substitute. Finally, the maximum-minimum normalization method is used to map the physical quantities to the [0,1] interval.
[0031] As a preferred approach, the normalized historical measured wind speed sequence is decomposed using a parameter-optimized VMD module to obtain several intrinsic mode function components, including: Modal number K With bandwidth factor α As a key parameter, the number of VMD modes has a decisive impact on the quality of VMD decomposition. To avoid the uncertainty of manual parameter selection, a genetic algorithm is introduced to evaluate the VMD mode number. K With bandwidth factor α Joint optimization is performed, and the objective function is: , In the formula, These are the components of each intrinsic mode function; The center frequency of each intrinsic mode function component; It is a pulse function; Modal number; For the first The modes obtained from the decomposition t Time is the independent variable; The exponent is a complex number, representing frequency shift and demodulation; The original signal; The energy entropy of the residual signal obtained by decomposing the historical measured wind speed sequence using VMD is used as the fitness function to measure the randomness and information concentration of the decomposed signal. Inconsistent interpretations are penalized through parameter boundary and integer constraint mechanisms. Finally, the optimal values of key parameters and the K intrinsic mode function components and decomposition residuals obtained after decomposition are output.
[0032] Specifically, the formula for calculating energy entropy is:
[0033] In the formula, This is the residual signal obtained after VMD decomposition. For the first The energy percentage of each sampling point.
[0034] Specifically, an improved genetic algorithm is used at the parameter search level. The genetic algorithm parameters are set as follows: population size is 30, maximum number of iterations is 25, crossover probability is 0.6, mutation probability is 0.2, tournament selection size is 5, and 2 elite individuals are retained to enhance search stability. An early stopping mechanism is also used to prevent invalid iterations.
[0035] Specifically, the final intrinsic mode function components are concatenated in the time dimension to form a multi-channel input feature vector. This feature vector serves as the input to the subsequent prediction model, characterizing the variation characteristics of the wind speed sequence at different frequency levels.
[0036] As a preferred approach, when the parameter-optimized VMD module decomposes the historical measured wind speed sequence, it constructs rolling standard deviation volatility characteristics with different window lengths for the historical measured wind speed sequence, making it easier for the prediction model to identify sudden wind speed changes and make reasonable predictions.
[0037] Specifically, the processed K intrinsic mode function components, the corresponding NWP wind speed sequence for the time period, and the predicted NWP future wind speed sequence for the time period are used to form a dataset, which is then divided into a training set, a test set, and a validation set in a ratio of 8:1:1.
[0038] Specifically, the dataset is divided into samples using a sliding window technique. The length of the sliding window is set, and data segments are extracted according to the time axis. The preferred length of the sliding window is 96 sampling points, and the sliding step size is 1.
[0039] S3. Use the training set and validation set to train the conditional residual gated recurrent neural network model Con-ResGRU to obtain the prediction model.
[0040] See Figure 2 The conditional residual gated recurrent neural network model Con-ResGRU is based on an encoder-decoder framework, and its specific structure includes an encoder, a multi-head self-attention mechanism, and a decoder arranged sequentially.
[0041] The Con-ResGRU takes the multi-channel feature vector formed by concatenating the intrinsic mode function components in the time dimension, and the NWP wind speed sequence from the same period as the input, and feeds it into the encoder after layer normalization of the input feature sequence. The encoder adopts the ResGRU (Residual Gated Recurrent Unit) model, which initially extracts the input temporal features, captures time-dependent information, and obtains the hidden representation of each historical moment. Among them, the intrinsic mode function components obtained by decomposing the historical measured wind speed sequence and the NWP wind speed sequence from the same period are simultaneously input into the model to help the model learn the systematic deviation between NWP and the actual wind speed.
[0042] The structure of the ResGRU model can be found in [link to ResGRU model]. Figure 3 Its structure includes an activation layer, a GRU layer, and a Dropout layer. The input features are activated by the ReLU function and then fed into the GRU to extract temporal features. The GRU output is then regularized by Dropout. The Dropout output is then concatenated with the input features through residuals, and finally, the output of ResGRU is obtained through Layer Normalization.
[0043] The ResGRU model used in the encoder of this embodiment preferably includes two vertically stacked GRUs, and the gating update process includes: , in, To update the door, To reset the door, , For the prediction period t Hidden states and candidate hidden states at each time step Representing the t- Decoding status at one time step, For the Sigmoid function, For Hadamard products, , , This represents the weight matrix and bias.
[0044] The multi-head self-attention mechanism is used to globally model the encoder output, thereby enhancing the expression of the long-range dependencies between different time steps of the historical measured wind speed sequence.
[0045] The multi-head self-attention mechanism in this embodiment is used to overcome the problem of the limited ability of traditional GRU to capture long-distance information in long sequence modeling. For example... Figure 4 As shown, this is the basic framework of the multi-head self-attention mechanism, which includes a linear mapping layer, a scaled dot product attention layer, a multi-head parallel attention layer, a splicing layer, and an output linear layer.
[0046] Specifically, the attention calculation process linearly maps the hidden state output by the encoder to the query. (Query), Key (Key) and value (Value), the single-head scaling dot product attention calculation is as follows:
[0047] In the formula, Represents the normalization function. Represents the key vector dimensionality; To model temporal dependencies in parallel from different subspaces, a multi-head self-attention mechanism is adopted:
[0048] In the formula, Represents concatenation operation. Represents the output projection matrix. The term represents an attention head, and its subscript indicates its number. Each attention head independently learns different temporal attention patterns, which helps the model to simultaneously capture trend changes and local fluctuations in the wind speed sequence.
[0049] As a preferred approach, to improve training stability and enhance feature representation capabilities, the multi-head self-attention mechanism in this embodiment incorporates residual connections and a feedforward network; the feedforward network is a two-layer fully connected network used to perform nonlinear feature reconstruction on the attention output, making the encoded features more suitable for future prediction.
[0050] Specifically, the multi-head self-attention mechanism performs global modeling of the encoder output, including:
[0051] in, The hidden state output by the encoder. for With multi-head self-attention output After summing the residuals, perform layer normalization. The obtained intermediate feature sequence, Feedforward network The nonlinear transformation output; The attention-enhanced feature sequence serves as the input context for the encoder and decoder.
[0052] The decoder is also based on the ResGRU model, which uses the output of the multi-head self-attention mechanism as its state input (i.e. Figure 2 (dec-input), using the future wind speed of the NWP as the conditional input (i.e. Figure 2 The prediction is incorporated into each step of the con-input, enabling the decoding layer to output future wind speed information that is closer to the original wind speed based on the NWP wind speed future information. Finally, the predicted wind speed for the future time step is obtained through mapping by the output layer.
[0053] The initial state of the decoder is the hidden state output by the encoder. Provides a temporal feature representation of historical wind speeds, and incorporates the NWP future wind speed at the corresponding time moment as the conditional input into the prediction at each time step, for example, in the first... t In the prediction at the nth time step, the decoder input not only includes the nth time step...t Decoding state of -1 time step (previous time step) It also explicitly introduces the NWP's future wind speed at the corresponding moment. As a conditional input, that is, the input to the decoder is The decoder outputs the prediction results for the prediction period step by step, realizing dynamic correction of the future wind speed of NWP.
[0054] As a preferred approach, to enhance the model's ability to correct systematic prediction biases, mitigate the vanishing gradient problem, achieve input-output mapping, and ensure module depth, this embodiment of Con-ResGRU introduces a global residual connection structure. Specifically, the decoder output outputs the predicted wind speed for the future time step through the global residual connection.
[0055] As a preferred approach, in order to further improve the model's ability to characterize the temporal variation of wind speed, enhance the model's ability to fit regions of sudden wind speed changes, and avoid the prediction results only fitting the numerical amplitude while ignoring the trend of wind speed changes, the prediction model training process uses a hybrid loss function weighted by the numerical error composed of the traditional mean square error (MSE) and the volatility loss term composed of the first-order difference error (Δ-MSE) to guide parameter updates.
[0056] Wherein, the hybrid loss function ,and: , , , In the formula, α and β are weighting coefficients. Mean square error, and The prediction model is respectively in the th t The, the t- Predicted wind speed output at one time step. and These are the corresponding actual wind speeds; H This is the total prediction step size. This is the volatility loss term for the first-order difference. and The first t The first-order difference between the actual wind speed and the predicted wind speed at each time step is used to balance the importance of numerical error and volatility loss constraints.
[0057] Specifically, the prediction results output by the prediction model are expressed as follows: , As a baseline forecast term This is the residual correction amount for the model output.
[0058] Specifically, the training set is input into the Con-ResGRU model for supervised training, and the validation set is used to monitor model performance and prevent overfitting. The learning rate is dynamically adjusted through the Adam optimizer until the model's classification accuracy on the validation set reaches the preset target. Finally, the model parameters with the highest accuracy and strongest generalization ability are saved.
[0059] The structure of the prediction model in this embodiment enables the model to focus on learning the deviation relationship between NWP and actual wind speed, thereby improving prediction accuracy and generalization ability.
[0060] S4. Input the test set into the prediction model to further verify the model performance and obtain the wind speed prediction results.
[0061] To verify and evaluate the performance of the prediction model in this embodiment, ablation experiments and comparative experiments using different models were conducted on the same dataset. MAE, RMSE, MAPE, and R... 2 The predictive effectiveness is evaluated using indicators such as [list of indicators]. The evaluation results are as follows: Figure 5 As shown, compared with models that do not introduce VMD decomposition (without VMD) or NWP conditional information (without Vol-loss) or other common models (ResGRU, LSTM, Transformer), the Con-ResGRU model scheme in this embodiment can achieve higher accuracy in wind speed prediction.
[0062] In summary, this invention utilizes both the multi-scale structural information of historical wind speed and the future prior information of NWP. Meanwhile, VMD decomposition helps to remove non-stationarity and multi-scale aliasing in historical sequences, enabling the model to better capture trends and fluctuations. Conditional decoding allows the model to better utilize future meteorological priors and closely approximate real wind speeds, thus providing a more reliable wind speed input basis for subsequent high-precision power prediction.
[0063] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A wind speed prediction method based on a conditional residual gated cyclic unit model, characterized in that, include: Obtain the historical measured wind speed sequence of the target wind turbine and the corresponding NWP wind speed sequence for the time period, and at the same time obtain the NWP future wind speed sequence for the prediction period; Data processing is performed on each wind speed sequence. The historical measured wind speed sequence is decomposed using the parameter-optimized VMD module to obtain several intrinsic mode function components. These components, together with the remaining wind speed sequences, constitute a dataset, which is then divided into a training set, a test set, and a validation set. The Con-ResGRU conditional residual gated recurrent neural network model is trained using the training and validation sets to obtain the prediction model; The Con-ResGRU comprises an encoder, a multi-head self-attention mechanism, and a decoder arranged sequentially. The encoder employs a ResGRU model, which uses a multi-channel feature vector formed by concatenating the intrinsic mode function components in the time dimension, along with the NWP wind speed sequence input, to initially extract the input temporal features and capture time-dependent information. The multi-head self-attention mechanism is used to globally model the encoder output, thereby enhancing the expression of long-range dependencies between different time steps of the historical measured wind speed sequence. The decoder, based on the ResGRU model, uses the output of the multi-head self-attention mechanism as the state input and the future NWP wind speed as the conditional input to incorporate into each prediction step, and obtains the predicted wind speed for the future time step through output layer mapping. The test set is input into the prediction model to obtain the wind speed prediction results.
2. The method according to claim 1, characterized in that, The VMD module with optimized parameters decomposes the historical measured wind speed sequence, including: Using genetic algorithms to analyze the key parameter of VMD: number of modes K With bandwidth factor α Joint optimization is performed, and the objective function is: ， In the formula, These are the components of each intrinsic mode function; The center frequency of each intrinsic mode function component; It is a pulse function; Modal number; For the first The modes obtained from the decomposition t Time is the independent variable; The exponent is a complex number, representing frequency shift and demodulation; The original signal; The energy entropy of the residual signal obtained by decomposing the historical measured wind speed sequence using VMD is used as the fitness function to measure the randomness and information concentration of the decomposed signal. Inconsistent interpretations are penalized through parameter boundary and integer constraint mechanisms. Finally, the optimal values of key parameters and the K intrinsic mode function components and decomposition residuals obtained after decomposition are output.
3. The method according to claim 2, characterized in that, When the parameter-optimized VMD module decomposes the historical measured wind speed sequence, it constructs rolling standard deviation volatility characteristics with different window lengths for the historical measured wind speed sequence, making it easier for the prediction model to identify sudden wind speed changes and make reasonable predictions.
4. The method according to claim 1, characterized in that, The encoder uses a ResGRU model consisting of two vertically stacked GRUs, and the gating update process includes: ， in, To update the door, To reset the door, , For the prediction period t Hidden states and candidate hidden states at each time step Representing the t- Decoding status at one time step, For the Sigmoid function, For Hadamard products, , , This represents the weight matrix and bias.
5. The method according to claim 4, characterized in that, The decoder outputs the first... t Hidden state at each time step As an initial state, at each time step, the decoder's input contains not only the hidden state from the previous time step. It also explicitly introduces the future wind speed of NWP at the corresponding time as a condition input, and then outputs the prediction results for the prediction period step by step to realize the dynamic correction of the future wind speed of NWP.
6. The method according to claim 1, characterized in that, The multi-head self-attention mechanism incorporates residual connections and a feedforward network; the feedforward network is a two-layer fully connected network used to perform nonlinear feature reconstruction on the attention output.
7. The method according to claim 6, characterized in that, The multi-head self-attention mechanism performs global modeling of the encoder output, including: ， in, The hidden state output by the encoder. for With multi-head self-attention output After summing the residuals, perform layer normalization. The obtained intermediate feature sequence, Feedforward network The nonlinear transformation output; The attention-enhanced feature sequence serves as the input context for the encoder and decoder.
8. The method according to claim 1, characterized in that, The Con-ResGRU is configured with a global residual connection, and the decoder output outputs the predicted wind speed for the future time step through the global residual connection.
9. The method according to claim 1 or 8, characterized in that, The prediction model uses a hybrid loss function that integrates mean squared error and volatility loss term to guide parameter updates; The hybrid loss function ,in: ， ， ， In the formula, α and β are weighting coefficients. Mean square error, and The prediction model is respectively in the th t The, the t- Predicted wind speed output at one time step. and These are the corresponding actual wind speeds; H This is the total prediction step size. This is the volatility loss term for the first-order difference. and The first t The first-order difference between the actual wind speed and the predicted wind speed at each time step is used to balance the importance of numerical error and volatility loss constraints.
10. The method according to claim 9, characterized in that, The prediction results output by the prediction model are expressed as follows: , As a baseline forecast term, This is the residual correction amount for the model output.