Wind speed prediction method based on deep learning
Through the wind speed prediction method of VMD decomposition and BiTCN/BiGRU parallel structure, the modeling problem of wind speed fluctuation and randomness is solved, and high-precision and stable wind speed prediction are achieved.
Patent Information
- Application Number
- CN202510465900.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-07-25
AI Technical Summary
Wind speed has strong volatility and randomness. The existing wind speed prediction methods are difficult to fully explore nonlinear relationships. The traditional fusion model has problems such as high computational complexity, error accumulation and poor information transmission. It is difficult to give full play to the advantages of each sub-model, affecting the stability and accuracy of the prediction.
Variable mode decomposition VMD is used to decompose wind speed data into multiple eigenmodal functions IMF, combining the bidirectional time convolution network BiTCN and the bidirectional gated cyclic network BiGRU parallel structure, and feature fusion is performed through the cross attention mechanism to generate wind speed prediction results.
It improves the stability and accuracy of wind speed prediction, reduces the difficulty of model learning, avoids error transmission and information loss, and enhances the modeling ability of complex changing wind speed data.
Smart Images
Figure CN120373551A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind speed prediction, and specifically provides a wind speed prediction method based on deep learning. Background Art
[0002] With the gradual depletion of traditional fossil energy and the increasingly severe global environmental problems, wind energy, as a renewable energy source, has received more and more attention globally. However, the wind speed has strong volatility and randomness characteristics, which pose great challenges to the stability of the power grid when wind power is integrated into the grid, affecting the wide application and utilization efficiency of wind energy.
[0003] Although many research progresses have been made in the field of wind speed prediction in recent years, there are still several key problems to be solved urgently. For example, 1. The wind speed has large volatility and the data itself has high complexity. Traditional analysis methods are difficult to fully explore the non-linear relationships therein, affecting the prediction accuracy. 2. Existing fusion models mostly rely on the way of cascading single models, ignoring the mutual influence between models. This not only increases the computational complexity and training difficulty, but also may cause problems such as error accumulation and poor information transmission. 3. When current fusion models combine the prediction results of multiple deep learning models, they mostly adopt simple methods such as splicing or weighted summation. When dealing with complex tasks, these methods have certain limitations and potential risks, and it is difficult to give full play to the advantages of each sub-model, thus affecting the stability and accuracy of the final prediction results. Therefore, there is an urgent need for a wind speed prediction technical solution based on deep learning to solve the above technical problems. Summary of the Invention
[0004] The purpose of the present invention is to provide a wind speed prediction method based on deep learning, which uses variational mode decomposition (VMD) to decompose wind speed data into multiple intrinsic mode functions (IMFs), and a parallel structure of a bidirectional temporal convolutional network (BiTCN) and a bidirectional gated recurrent network (BiGRN) is used to extract the short-term and long-term features of the intrinsic mode functions respectively, and then the long-term and short-term features are fused with weights to obtain the wind speed prediction result. This is to solve the problems raised in the above background art.
[0005] To solve the above technical problems, the present invention provides a wind speed prediction method based on deep learning, including:
[0006] Step 1: Use linear interpolation method to fill in the missing values of the original wind speed data.
[0007] Set the wind speed as in there are two data points and . When there is a missing value for the sampling point between the two data points, the missing value is calculated by the formula: ; In the formula, and are the corresponding values of the sampling points and respectively, and is the missing value corresponding to the sampling point .
[0008] Quickly fill in the missing data to avoid the deviation of model training caused by data gaps. Maintain the local linear trend of the data in the time series, without introducing high-frequency noise, and is suitable for subsequent decomposition of stationary signals.
[0009] Step 2: Calculate the central frequency of each intrinsic mode function IMF.
[0010] The calculation formula for the central frequency is: ; In the formula, is the central frequency of the th IMF. is the representation of the th IMF in the frequency domain, that is, the frequency domain signal obtained through Fourier transform and other transformations. is the frequency variable, covering all frequency components during integration. is the energy of the IMF component at the frequency .
[0011] The numerator part of the integral operation represents the weighted integral of the frequency and its corresponding IMF energy, reflecting the "center of gravity" of the energy distribution. The denominator part represents the total energy of the IMF.
[0012] The central frequency reflects the main fluctuation characteristics of the IMF component. High-frequency IMF corresponds to short-term fluctuations, and low-frequency IMF corresponds to long-term trends. It provides a quantitative basis for selecting the optimal K value later.
[0013] The central frequency is essentially a representative frequency obtained through energy weighted average. Quantify the main frequency characteristics of the IMF component through the way of energy weighted average, providing a basis for determining the number of modes of VMD decomposition in the subsequent steps.
[0014] Step 3: Determine the number of modes K of variational mode decomposition VMD through the central frequency observation method.
[0015] By setting different K values and observing the change of the central frequency, select the K value when the same-frequency mode first appears in the central frequency as the optimal decomposition mode number.
[0016] Adopt an adaptive decomposition method to avoid over-decomposition or under-decomposition of the signal and improve the decomposition quality.
[0017] Step 4: Use VMD to decompose the wind speed data into K IMFs with different central frequencies.
[0018] Decompose the wind speed data obtained in Step 1 into IMFs with different central frequencies. After the decomposition is completed, save each extracted IMF component and further combine them into new features for subsequent modeling analysis.
[0019] Separate the non-stationary wind speed data into sub-signals of different scales to reduce the model learning difficulty. The IMF components can be directly used as input features for the subsequent model, providing clear short-term and long-term dynamic information.
[0020] Step 5: Normalize the decomposed IMFs.
[0021] The normalization formula for IMFs is: ; where is the th normalized IMF, is the th IMF, is the mean of the th IMF, is the th standard deviation of the IMF.
[0022] Unify the numerical ranges of different IMFs to prevent the model from being biased towards certain components due to input scale differences. The normalized data better conforms to the input distribution of the neural network, improving the training efficiency.
[0023] Step 6: Input the normalized IMFs into the parallel Bidirectional Temporal Convolutional Network (BiTCN) and Bidirectional Gated Recurrent Unit (BiGRU) to extract short-term features and long-term features respectively.
[0024] BiTCN extracts the forward and backward short-term features of the sequence through bidirectional dilated causal convolution, and BiGRU captures the long-term dependency dynamic features through a bidirectional recurrent network.
[0025] Integrate local fine-grained features and global temporal dependencies to enhance the model's representation ability for the multi-scale characteristics of wind speed. The parallel structure reduces the dependency problem between the front and back modules in the serial model.
[0026] Step 7: Use the cross-attention mechanism to take the short-term features as the query Query, the long-term features as the key Key and the value Value, and perform dynamic weighted fusion to obtain the fused features.
[0027] Specific implementation steps:
[0028] First, adopt the cross-attention mechanism for the short-term features As the query, the long-term feature As the key Key and value Value.
[0029] Secondly, by calculating and The similarity between them is as follows: ; ; ; In the formula, , , Are the values after linear transformation respectively, , , Are the learnable weight parameters respectively, Is the feature dimension, Is the transformed feature dimension.
[0030] Finally, for and Perform dynamic weighted fusion to obtain the fused feature : ; In the formula, Is the weight normalization function.
[0031] Represents the transpose of the key matrix, that is, the positions of the row and column elements of the matrix are interchanged. By the transpose operation, the dimensions of the matrices and Are aligned to calculate the similarity between the two.
[0032] Dynamically adjust the attention weight to the long-term feature according to the short-term change, and enhance the attention to the recent pattern when the wind speed suddenly changes. Combine short-term details with long-term trends to reduce the information one-sidedness of a single feature.
[0033] Step 8: Input the fused feature into the fully connected layer for mapping to generate the prediction result.
[0034] The fused feature Is mapped to the output dimension through the fully connected layer to obtain the wind speed prediction result : ; In the formula, Is the weight matrix, Is the bias vector.
[0035] Map the high-dimensional fused feature to a specific wind speed value to achieve end-to-end prediction. The parameters of the fully connected layer can be optimized through training to directly associate the feature with the prediction target.
[0036] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0037] 1. Deep collaboration between signal decomposition and deep learning: The wind speed data is decomposed into high-frequency and low-frequency components through variational mode decomposition (VMD), which not only retains the short-term fluctuation details of the wind speed but also extracts the long-term evolution trend. By combining subsequent deep learning methods to model different components separately, the limitation of traditional analysis methods in parsing complex non-linear relationships is broken through, and the problem of difficult modeling of the volatility and randomness of wind speed data is solved.
[0038] 2. Parallel complementarity of dual models and anti-error accumulation: A parallel structure of bidirectional temporal convolutional network (BiTCN) and bidirectional gated recurrent unit (BiGRU) is adopted to extract short-term local features and long-term dependence features respectively, avoiding the error transmission problem caused by front-back dependence in traditional serial models. The independence between models realizes complementary advantages, avoiding the learning bias and information loss of single models, and improving the modeling ability for complex changing wind speed data.
[0039] 3. Dynamic interactive feature fusion mechanism: Using the cross-attention mechanism, the short-term features are used as driving signals to dynamically adjust the fusion weights of long-term features, replacing the traditional static fusion methods such as weighted summation or splicing. This mechanism can adapt to complex scenarios of wind speed changes, fully integrate local details and global trend information, avoid traditional splicing or weighted summation operations, reduce implicit feature conflicts, and improve prediction accuracy and robustness.
[0040] 4. Adaptive optimization of modal decomposition parameters: The number of modal decompositions K is selected based on the central frequency distribution law of IMF components. Through energy analysis, the randomness of artificially setting parameters is avoided, and the phenomenon of over-decomposition or under-decomposition that may occur in conventional VMD is suppressed, ensuring that the features input into the deep learning model have clear physical meanings and separability. Brief Description of the Drawings
[0041] The drawings are used to provide further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:
[0042] Figure 1 is a schematic flowchart of a wind speed prediction method based on deep learning according to the present invention;
[0043] Figure 2 is a schematic diagram of the VMD decomposition wind speed data result of a wind speed prediction method based on deep learning according to the present invention;
[0044] Figure 3 is a schematic diagram of the structure of a bidirectional dilated causal convolutional network of a wind speed prediction method based on deep learning according to the present invention;
[0045] Figure 4 It is a schematic diagram of the bidirectional gated recurrent unit network structure of a wind speed prediction method based on deep learning according to the present invention. Detailed implementation manners
[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0047] Please refer to Figure 1 , the present invention provides a wind speed prediction method based on deep learning, including:
[0048] Step 1: Use linear interpolation to fill in the missing values of the original wind speed data.
[0049] Set the wind speed as , and there are two data points and in . When there is a missing value at the sampling point between the two data points, calculate the missing value through the formula: ; In the formula, and are the corresponding values of the sampling points and respectively, and is the missing value corresponding to the sampling point .
[0050] Quickly fill in the missing data, avoid the training deviation of the model caused by the data gap. Maintain the local linear trend of the data in the time series, and will not introduce high-frequency noise, which is suitable for subsequent stationary signal decomposition.
[0051] Step 2: Calculate the central frequency of each intrinsic mode function IMF.
[0052] The calculation formula of the central frequency is: ; In the formula, is the central frequency of the th IMF. is the representation of the th IMF in the frequency domain, that is, the frequency domain signal obtained through Fourier transform and other transformations. is the frequency variable, and all frequency components are covered during integration. is the IMF component at frequency The energy at (power spectral density).
[0053] The numerator of the integral operation represents the weighted integral of the frequency and its corresponding IMF energy, reflecting the "center of gravity" of the energy distribution. The denominator represents the total energy of the IMF.
[0054] The center frequency reflects the main fluctuation characteristics of the IMF component. High-frequency IMF corresponds to short-term fluctuations, and low-frequency IMF corresponds to long-term trends. It provides a quantitative basis for selecting the optimal K value later.
[0055] The central frequency is essentially a representative frequency obtained by energy-weighted averaging. The main frequency characteristics of the IMF components are quantified by energy-weighted averaging, providing a basis for determining the number of modes of VMD decomposition in the subsequent steps.
[0056] Step 3: Determine the mode number K of variational mode decomposition VMD by the center frequency observation method.
[0057] By setting different K values and observing the changes in the center frequency, the K value when the center frequency first appears with the same frequency mode is selected as the optimal number of decomposed modes.
[0058] Adaptive decomposition is used to avoid over-decomposition (interfering with subsequent modeling) or under-decomposition (feature mixing) of the signal, thereby improving the decomposition quality.
[0059] Step 4: Please refer to Figure 2 , VMD is used to decompose the wind speed data into K IMFs with different center frequencies.
[0060] The wind speed data obtained in step 1 is decomposed into IMFs with different center frequencies. After the decomposition is completed, the extracted IMF components are saved and further combined into new features for subsequent modeling and analysis.
[0061] Separate non-stationary wind speed data into sub-signals of different scales (such as high-frequency noise and low-frequency trends) to reduce the difficulty of model learning. IMF components can be directly used as input features of subsequent models to provide clear short-term and long-term dynamic information.
[0062] Step 5: Normalize the decomposed IMF.
[0063] The normalized IMF formula is: ; In the formula, It is Normalized IMF, It is IMF, It is The mean value of an IMF, is the standard deviation of the
[0064] Unify the value ranges of different IMFs to prevent the model from being biased towards certain components due to differences in input scales. Normalized data better conforms to the input distribution of the neural network and improves training efficiency.
[0065] Step 6: Input the normalized IMFs into the parallel Bidirectional Temporal Convolutional Network (BiTCN) and Bidirectional Gated Recurrent Unit (BiGRU) to extract short-term features and long-term features respectively.
[0066] Please refer to Figure 3 , BiTCN extracts the forward and backward short-term features of the sequence through bidirectional dilated causal convolution.
[0067] Please refer to Figure 4 , BiGRU captures the long-term dependence dynamic features through a bidirectional recurrent network.
[0068] Integrate local fine-grained features and global temporal dependencies to enhance the model's ability to represent the multi-scale characteristics of wind speed. The parallel structure reduces the dependency problem between the front and back modules in a cascaded model.
[0069] Step 7: Use the cross-attention mechanism to take the short-term features as the query Query, the long-term features as the key Key and value Value, and perform dynamic weighted fusion to obtain the fused features.
[0070] Specific implementation steps:
[0071] First, adopt the cross-attention mechanism to take the short-term features as the query Query, and the long-term features as the key Key and value Value.
[0072] Secondly, by calculating the and similarity between them, the formula is as follows: ; ; ; In the formula, , , are the values after linear transformation respectively, , , are the learnable weight parameters respectively, is the feature dimension, is the transformed feature dimension.
[0073] Finally, for and perform dynamic weighted fusion to obtain fused features : ; In the formula, is the weight normalization function.
[0074] represents the transpose of the key matrix, that is, the positions of the row and column elements of the matrix are interchanged. Through the transpose operation, the matrices (Query) and (Key) are dimensionally aligned, so as to calculate the similarity (dot product) between the two.
[0075] Dynamically adjust the attention weight to the long-term features according to the short-term changes, and enhance the attention to the recent patterns when the wind speed changes suddenly. Combine short-term details with long-term trends to reduce the information one-sidedness of single features.
[0076] Step 8: Input the fused features into the fully connected layer for mapping to generate prediction results.
[0077] The fused features are mapped to the output dimension through the fully connected layer to obtain the wind speed prediction result : ; In the formula, is the weight matrix, is the bias vector.
[0078] Map the high-dimensional fused features into specific wind speed values to achieve end-to-end prediction. The parameters of the fully connected layer can be optimized through training, directly associating features with prediction targets.
[0079] Example 1:
[0080] In this example, the central frequency observation method is used to select the value of K. By setting different values of K, the central frequencies of each IMF component are obtained, as shown in Table 1:
[0081] Table 1
[0082]
[0083] Table 1 shows that when K = 6, modes with similar central frequencies begin to appear, resulting in over-decomposition. Therefore, when the number of IMFs K is selected as 6, the VMD decomposition effect is the best.
[0084] Example 2:
[0085] In this example, RMSE, MAE, MAPE and As the key indicators for evaluating the performance of the model, the smaller the values of RMSE, MAE, and MAPE, the better the model's prediction effect, and the closer the predicted value is to the actual value. While the closer it is to 1, the higher the fitting degree of the model and the better the prediction effect. Their calculation formulas are as follows: ; ; ; ; Among them, is the actual value, is the predicted value, is the mean value of the actual values, is the number of sampling points.
[0086] The prediction results were compared with the VMD-BiTCN and VMD-BiGRU deep learning models. The dataset was sampled at hourly intervals. To meet the prediction requirements for different time scales, the prediction step sizes were set to 4 steps, 8 steps, 16 steps, and 24 steps, corresponding to the prediction results for the next 4 hours, 8 hours, 16 hours, and 24 hours respectively.
[0087] The evaluation index results of the prediction results of each model at different prediction step sizes are shown in Table 2.
[0088] Table 2
[0089]
[0090] As can be seen from the above table, among the prediction results of the comparative tests, the evaluation indexes of the method of the present invention are the best. For example, in the 4-step prediction, the RMSE of the method of the present invention is 0.076, the MAE is 0.059, and the MAPE is 0.352%, is 0.988, which is better than other deep learning models. In the case of other step numbers, the evaluation indexes of the method of the present invention are also better than other deep learning models.
[0091] In summary, the method of the present invention performs excellently at each prediction step size. Especially in the prediction of long time scales, it maintains high accuracy and stability, and is better than other comparative models.
[0092] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variation thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.
[0093] Finally, it should be noted that the above are only preferred embodiments of the present invention and are 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 perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A wind speed prediction method based on deep learning, characterized in that: The method includes: Step 1: Fill in the missing values of the original wind speed data using linear interpolation; Step 2: Calculate the central frequency of each Intrinsic Mode Function (IMF); Step 3: Determine the number of modes K of Variational Mode Decomposition (VMD) by observing the central frequency; Step 4: Use VMD to decompose the wind speed data into K IMFs with different central frequencies; Step 5: Normalize the decomposed IMFs; Step 6: Input the normalized IMFs into the parallel Bidirectional Temporal Convolutional Network (BiTCN) and Bidirectional Gated Recurrent Unit (BiGRU) to extract short-term features and long-term features respectively; Step 7: Use the cross-attention mechanism to take the short-term features as the query Query, and the long-term features as the key Key and value Value, and perform dynamic weighted fusion to obtain the fused features; Step 8: Input the fused features into the fully connected layer for mapping to generate the prediction result.
2. The wind speed prediction method based on deep learning according to claim 1, wherein: In step 1, set the wind speed to , and there are two data points in and ; when there is a missing value for the sampling point between the two data points, calculate the missing value through the formula: ； In the formula, and are the corresponding values of the sampling points and respectively, is the missing value corresponding to the sampling point respectively.
3. The wind speed prediction method based on deep learning according to claim 1, wherein: In Step 2, the calculation formula for the central frequency is: ； Wherein, is the center frequency of the th IMF; is the representation of the th IMF in the frequency domain, that is, the frequency domain signal obtained through Fourier transform or other transforms; is the frequency variable.
4. A wind speed prediction method based on deep learning according to claim 1, characterized in that: In Step 3, by setting different K values and observing the change of the central frequency, select the K value when the central frequency first appears in the same-frequency mode as the optimal decomposition mode number.
5. A wind speed prediction method based on deep learning according to claim 1, characterized in that: In Step 4, decompose the wind speed data obtained in Step 1 into IMFs with different central frequencies; after decomposition, save each extracted IMF component and further combine them into new features for subsequent modeling analysis.
6. The wind speed prediction method based on deep learning according to claim 1, wherein: In Step 5, the formula for normalizing the IMF is: ； wherein, is the th normalized IMF, is the th IMF, is the mean value of the th IMF, is the standard deviation of the th IMF.
7. A wind speed prediction method based on deep learning according to claim 1, characterized in that: In Step 6, the BiTCN extracts the forward and backward short-term features of the sequence through bidirectional dilated causal convolution, and the BiGRU captures the long-term dependence dynamic features through the bidirectional recurrent network.
8. A wind speed prediction method based on deep learning according to claim 1, characterized in that: In Step 7, the specific implementation steps: First, the cross-attention mechanism is used to take the short-term features as the query Query, and the long-term features as the key Key and value Value; Secondly, by calculating and the similarity between them, the formula is as follows: ； ； ； Wherein, , , are respectively the values after linear transformation, , , are respectively the learnable weight parameters, is the feature dimension, is the feature dimension after transformation; Finally, perform dynamic weighted fusion on and to obtain the fused feature : ； In the formula, is the weight normalization function.
9. A wind speed prediction method based on deep learning according to claim 8, characterized in that: In step 8, the fused feature is mapped to the output dimension through a fully connected layer to obtain the wind speed prediction result : ； In the formula, is the weight matrix, is the bias vector.