Frequency characteristic-oriented wind power prediction method and device and storage medium

Abstract

The invention provides a frequency characteristic-oriented wind power prediction method and device and a storage medium, and belongs to the technical field of new energy and intelligent power systems. According to the wind power prediction method for the frequency characteristics provided by the embodiment of the invention, a hybrid model fusing CEEMDAN decomposition, VMD denoising, an SE evaluation mechanism and Transform-GRU double-structure modeling is provided according to the specific scheme, and high-frequency noise interference is effectively reduced when a wind turbine generator set outputs a non-stationary sequence. Moreover, the frequency characteristics are fully utilized, and the differential learning mechanism design is adopted according to the frequency characteristics and complexity differences of the decomposition modal units, so that the feature extraction and output precision and the time sequence dynamic capture capability are improved. Differentiated modeling and efficient prediction of different frequency modal units can be achieved, the precision and stability of ultra-short-term prediction of wind power are effectively improved, and the method is suitable for wind power grid-connected scheduling and system operation optimization requirements under complex meteorological conditions.

Description

Technical Field

[0001] This invention relates to the field of new energy and smart power system technology, and in particular to a wind power prediction method, device and storage medium oriented towards frequency characteristics. Background Technology

[0002] Among various forms of renewable energy, wind power has become one of the fastest-growing new energy sources due to its high efficiency, cleanliness, and economy. However, the utilization of wind power still faces many challenges. The output characteristics of wind power are significantly affected by factors such as geographical location and weather changes, exhibiting strong volatility, randomness, and intermittency, which poses stability risks to large-scale grid-connected operation. Therefore, wind power forecasting technology is of great significance in improving grid dispatch efficiency, optimizing the coordinated operation of wind and thermal power, and enhancing the absorption capacity of renewable energy, and has become one of the key research directions in the field of wind power grid connection.

[0003] In related technologies, wind power forecasting can be categorized into ultra-short-term forecasting, short-term forecasting, medium-term forecasting, and long-term forecasting based on different prediction time scales. Among these, ultra-short-term forecasting plays a crucial role in assisting wind turbine operation control and grid load tracking. Addressing the nonlinear and non-stationary characteristics of wind power forecasting, the advantages of recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs) in time series modeling have become increasingly apparent with the development of deep learning, leading to their widespread application in wind power forecasting. However, a single model struggles to simultaneously process information from different frequency components, limiting modeling performance. Therefore, the "decomposition and integration" prediction framework—signal decomposition-modeling-integration—has become a research hotspot. This method decomposes complex wind power sequences into several modal units with distinct frequency characteristics through signal decomposition, models each modality individually, and then integrates the prediction results, effectively mitigating nonlinear interference and improving prediction accuracy.

[0004] In recent years, scholars have proposed various decomposition-fusion prediction frameworks based on the above strategies, such as CEEMDAN-CNN-LSTM, CEEMDAN-BiLSTM, and CEEMDAN-Transformer models, which have improved prediction accuracy to some extent. However, most methods adopt a unified modeling structure for all decomposed modal units, failing to fully consider the differences in their frequency characteristics and modeling complexity. Furthermore, the models still suffer from insufficient coupling in handling high-frequency nonlinear features and low-frequency time-dependent features. This results in poor dynamic structure capture capability for ultra-short-term fluctuations in wind power based on existing technologies, leading to insufficient prediction accuracy and robustness. Summary of the Invention

[0005] This invention provides a wind power prediction method, device, and storage medium based on frequency characteristics. By constructing a hybrid model, it achieves differentiated modeling and efficient prediction of different frequency mode units, effectively improving the accuracy and stability of wind power prediction. It is suitable for wind power grid-connected scheduling and system operation optimization needs under complex meteorological conditions. The technical solution is as follows: In a first aspect, embodiments of the present invention provide a wind power prediction method oriented towards frequency characteristics, comprising: Step 1: Obtain the original wind power output sequence of the wind turbine. Step 2: The original wind power sequence is decomposed using the adaptive noise complete set empirical mode decomposition method to obtain multiple intrinsic mode functions; Step 3: Use the variational mode decomposition algorithm to perform a secondary decomposition on the multiple intrinsic mode functions to decompose them into multiple modal units with finite bandwidth. Step 4: Use the sample entropy complexity evaluation method to evaluate and classify the complexity of the multiple modal units, so as to divide the multiple modal units into two categories: high-frequency modal units and low-frequency modal units; Step 5: Use the Transformer model to model and predict the high-frequency modal units, and output the first modal unit output response; use the GRU model to predict and model the low-frequency modal units, and output the second modal unit output response. Step six: Based on the output responses of the first and second modal units, integrate and reconstruct the final wind power prediction results.

[0006] Optionally, step one includes: The original wind turbine output sequence corresponding to the historical sampling time of the wind farm is obtained. Missing values ​​are processed, outliers are removed and normalized to obtain the original wind power output sequence.

[0007] Optionally, step two includes: To the original wind power sequence Add to White noise with a mean of 0 ,generate A new wind power sequence ; The empirical mode decomposition method was used to analyze each group of wind power sequences. Decompose to obtain Each component This leads to the first eigenmode function. and abnormal signal components ; To the layer abnormal signal component Add adaptive noise Furthermore, the empirical mode decomposition (EMD) method is used for further decomposition and mean processing to obtain the second intrinsic mode function. and new residual fluctuation signals ; Recursion to find the first Individual eigenmode functions and the A residual signal of fluctuation ; Repeat the above steps until the first... A residual signal of fluctuation If the function is monotonic, the adaptive noise complete set empirical mode decomposition method terminates.

[0008] Optionally, step four includes: In a given embedding dimension and similar threshold Under the condition of reconstructing the embedding space of the modal unit, the statistical dimension is then calculated. and The logarithm of the matching vector is used to calculate the logarithmic ratio of the multiple modal units. The average value of the logarithmic ratios corresponding to the multiple modal units is set as the evaluation threshold. Modal units with logarithmic ratios higher than the evaluation threshold are classified as high-frequency modal units, and modal units with logarithmic ratios lower than the evaluation threshold are classified as low-frequency modal units.

[0009] Optionally, step five includes: The encoder of the Transformer model is used to convert the high-frequency modal units into high-dimensional feature representations to capture short-term fluctuations and long-term dependencies. The decoder of the Transformer model generates the first modal unit output response corresponding to each of the high-frequency modal units based on the short-term fluctuations and the long-term dependencies.

[0010] Optionally, step six includes: The output responses of the first and second modal units are denormalized, and the predicted data after denormalization are summed to obtain and output the wind power prediction result.

[0011] Optionally, the original wind turbine output sequence is the actual wind power data of the wind turbines in the wind farm.

[0012] In a second aspect, the present invention provides a wind power prediction device, comprising: The preprocessing module is used to obtain the raw wind power output sequence of the wind turbine. The preliminary decomposition module is used to decompose the original wind power sequence using the adaptive noise complete set empirical mode decomposition method to obtain multiple intrinsic mode functions; The secondary decomposition module is used to perform secondary decomposition on the multiple intrinsic mode functions using a variational mode decomposition algorithm, so as to decompose them into multiple modal units with finite bandwidth. The evaluation module is used to evaluate and classify the complexity of multiple modal units using the sample entropy complexity evaluation method, so as to divide the multiple modal units into two categories: high-frequency modal units and low-frequency modal units. The modeling and prediction module is used to model and predict the high-frequency modal units using the Transformer model and output the output response of the first modal unit; and to predict and model the low-frequency modal units using the GRU model and output the output response of the second modal unit. The result integration module is used to integrate and reconstruct the wind power prediction result based on the output response of the first modal unit and the output response of the second modal unit.

[0013] Thirdly, the present invention also provides a wind power prediction device, the wind power prediction device comprising: processor; Memory used to store instructions that can be executed by the processor; The processor is configured to execute the frequency-oriented wind power prediction method described in the first aspect above.

[0014] Fourthly, the present invention also provides a storage medium storing computer instructions thereon, which, when executed by a processor, implement the frequency-oriented wind power prediction method described in the first aspect above.

[0015] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: The wind power prediction method based on frequency characteristics provided in this invention proposes a hybrid model that integrates CEEMDAN decomposition, VMD denoising, SE evaluation mechanism, and Transformer-GRU dual-structure modeling. This model effectively reduces high-frequency noise interference when dealing with non-stationary wind turbine output sequences. Furthermore, it fully utilizes frequency characteristics, employing a differentiated learning mechanism design based on the frequency characteristics and complexity differences of the decomposed modal units to improve feature extraction, output accuracy, and time-series dynamic capture capabilities. This enables differentiated modeling and efficient prediction of modal units at different frequencies, effectively improving the accuracy and stability of ultra-short-term wind power prediction, and is suitable for wind power grid-connected scheduling and system operation optimization needs under complex meteorological conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a wind power prediction method based on frequency characteristics provided in an embodiment of the present invention; Figure 2 This is a control structure block diagram of a wind power prediction device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the control structure of another wind power prediction device provided in an embodiment of the present invention; Figure 4 This is a wind power prediction framework diagram based on CEEMDAN-VMD-SE-Transformer-GRU provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of wind energy data for La Haute Borne, France in 2017, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of high-frequency modal units after SE classification of modal units provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of low-frequency modal units after SE classification of modal units provided in an embodiment of the present invention; Figure 8 This is a comparison chart of the wind power prediction curve and the actual wind power curve provided in the embodiments of the present invention.

[0018] In the picture: 1-Preprocessing module; 2-Preliminary decomposition module; 3-Secondary decomposition module; 4-Evaluation module; 5-Modeling and prediction module; 6-Result integration module. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart of a wind power prediction method based on frequency characteristics provided in an embodiment of the present invention. Figure 4 This is a wind power prediction framework diagram based on CEEMDAN-VMD-SE-Transformer-GRU provided in an embodiment of the present invention. (See diagram below.) Figure 1 combine Figure 4As shown, through practice, the inventors have provided a wind power prediction method based on frequency characteristics, which includes: S1, obtain the original wind power output sequence of the wind turbine.

[0021] Specifically, in this step, when subsequent wind power prediction of wind turbines is required, the original wind turbine output sequence corresponding to the historical sampling time of the wind farm is obtained. The original wind turbine output sequence is the actual wind power data of the wind turbines in the wind farm. By handling missing values, removing outliers, and normalizing the data, a clean and stable input is provided for subsequent model training to obtain the original wind power output sequence.

[0022] S2. The original wind power sequence is decomposed using the CEEMDAN method of adaptive noise complete set empirical mode decomposition to obtain multiple intrinsic mode functions.

[0023] CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) is a signal processing technique that decomposes nonlinear, non-stationary data into multiple more stable and stationary intrinsic mode functions (IMFs). Specifically, this process includes: S21, to the original wind power sequence Add to White noise with a mean of 0 ,generate A new wind power sequence : (1) S22, the Empirical Mode Decomposition (EMD) method is used to process each wind power sequence. Decompose to obtain Each component This leads to the first eigenmode function. and abnormal signal components : (2) (3) S23, layer anomaly signal component Add adaptive noise Furthermore, the empirical mode decomposition (EMD) method is used for further decomposition and mean processing to obtain the second intrinsic mode function. and new residual fluctuation signals : (4) (5) In the formula: This indicates the first generation of the signal after EMD decomposition. Quantity, This is the noise weighting coefficient.

[0024] S24, recursively to find the... Individual eigenmode functions and the A residual signal of fluctuation : (6) (7) S25, repeat the above steps until the... A residual signal of fluctuation Since it is a monotonic function, the CEEMDAN algorithm terminates, and the final fluctuation residue can be calculated according to equation (8): (8) S3 utilizes the Variational Mode Decomposition (VMD) algorithm to perform secondary decomposition on multiple intrinsic mode functions, thereby decomposing them into multiple modal units with finite bandwidth.

[0025] Specifically, in this step, in order to further reduce the decomposition of CEEMDAN To mitigate noise interference, the VMD algorithm is introduced for secondary decomposition, reducing frequency aliasing and nonlinear characteristics of the data, thereby improving the predictive performance of subsequent models. VMD is based on a variational model, decomposing the signal into K modal components with finite bandwidth. It satisfies the following approximation relationship: (9) S4 utilizes the Sample Entropy Complexity Evaluation (SE) method to evaluate and classify the complexity of multiple modal units, thereby dividing them into two categories: high-frequency modal units and low-frequency modal units.

[0026] Although modal units have been partitioned in the frequency domain, different modes may still exhibit significant differences in dynamic complexity. To further differentiate the learning mechanisms for each mode, this invention introduces Spectral Entropy (SE) to evaluate and classify the complexity of all modal units. SE is an indicator of time series complexity, an improvement upon approximate entropy. Generally, the larger the SE, the more complex the internal structure of the sequence, the stronger the volatility, and the higher the prediction difficulty.

[0027] Specifically, the steps include: S41, given the embedding dimension and similar threshold Under the condition of reconstructing the embedding space of the modal unit, the statistical dimension is then calculated. and The logarithm of the matching vectors is used to calculate the logarithmic ratio of multiple modal units. A modal unit is defined as... Then its SE representation is as follows: (10) in, and The embedding dimensions are respectively and The number of vector pairs that satisfy the matching criteria can be expressed as: (11) (12) in, express The number of similar vectors.

[0028] S42, set the average value of the logarithmic ratios corresponding to multiple modal units as the evaluation threshold, classify modal units with logarithmic ratios higher than the evaluation threshold as high-frequency modal units, and classify modal units with logarithmic ratios lower than the evaluation threshold as low-frequency modal units.

[0029] Specifically, based on the calculation results, this invention sets the evaluation threshold as the average value of SE, dividing all modal units into two categories: high-frequency modal units and low-frequency modal units, to facilitate the subsequent adoption of targeted modeling methods. Furthermore, through the above formula, we found that SE is related to the embedding dimension. and similar threshold Relevant. In practical applications, it is generally set... It is 2 or 3. It is 0.1 to 0.25 times the standard deviation. To balance computational stability and resolution, this invention sets... , This parameter combination performs well in nonlinear time series analysis tasks.

[0030] S5 uses the Transformer model to model and predict the high-frequency modal unit and outputs the output response of the first modal unit; it uses the GRU model to predict and model the low-frequency modal unit and outputs the output response of the second modal unit.

[0031] The high-frequency and low-frequency modal units of the classification are normalized separately and then input into different prediction models. The high-frequency modal units are processed using the Transformer model to extract long-term dependencies and complex dynamic patterns; the low-frequency sequences are input into the GRU (Gated Recurrent Unit) model, taking advantage of its simple structure and efficient training for modeling and prediction.

[0032] Specifically, for high-frequency modal unit modeling, in this study, the high-frequency modal units classified by SE are input into Transformer. Considering that the input data consists of multiple sets of decomposed modal units and the main objective is time-dependent modeling, the encoder part of Transformer is used as a feature extraction structure, while the decoder part is responsible for mapping the features extracted by the encoder to the target prediction sequence, and finally realizing the prediction of the wind turbine output.

[0033] The Transformer encoder aims to transform the decomposed wind power output modal units into high-dimensional feature representations, capturing short-term fluctuations and long-term dependencies. Each input modal unit is represented as... Where T is the backtracking window length, first projected onto Dimensional space: (13) in, and These are the weight matrix and the bias term.

[0034] Since the Transformer itself does not have the ability to perceive temporal information, a positional encoding based on sine and cosine functions is introduced and added element-wise to the input embedding vector to explicitly inject temporal order information. The calculation method is as follows: (14) (15) in, Indicating the time step position, the input embedding is subsequently augmented to PE is the location code, i is the dimension index, and d is the embedding dimension.

[0035] The core of the encoder is the self-attention mechanism, which calculates the weighted relationship between the query (Q), key (K), and value (V) to determine the similarity between positions in the input sequence and capture global dependencies. (16) in, Let be the dimension of the key vector. Pay attention to the number of heads. The function is used to normalize the attention weights, thereby achieving weighted fusion of information. To enhance the model's expressive power in different subspaces, a multi-head attention mechanism is introduced, which computes attention in parallel across multiple subspaces. Then perform splicing and linear transformation: (17) (18) in, , , For the The projection matrix of the head, To output the mapping matrix, This is for splicing operations.

[0036] After the self-attention output, the representation at each position is fed into a feedforward neural network (FFN). This module has a two-layer fully connected neural network structure and non-linear transformation capabilities. (19) in, For input features, , This is the weight matrix. , For bias terms, For the hidden layer dimension.

[0037] To mitigate overfitting and ensure training stability, oscillatory residual connections and layer normalization are applied after each sublayer: (20) The encoder contains The same layer, the final output As a representation of the upper and lower information of the wind power mode unit, it reflects its temporal dynamics.

[0038] The decoder utilizes the encoder output and models the time dependency of the target sequence to generate the wind power output response. Given the target sequence... . To predict the step size, similarly, it is first embedded in the d-dimensional space according to equation (13), and the position encoding is performed according to equations (14) and (15) to form the decoder input. .

[0039] The decoder contains two attention sublayers, and the multi-head attention is calculated according to equations (17) and (18). Furthermore, a masked self-attention mechanism processes the target sequence: (twenty one) in, , , , The mask matrix ensures that each position focuses only on the preceding time step, thus maintaining the autoregressive nature of wind power output prediction.

[0040] Furthermore, the encoder-decoder attention mechanism integrates the encoder output. With the target sequence: (twenty two) in, , For self-attention output, , The decoder focuses on the relevant temporal features of the input modal units, aligning them with the prediction task.

[0041] The attention output is processed using FFN according to Equation (19), followed by ripple residual connections and layer normalization according to Equation (20). The decoder contains... Each layer, the final output The output response is mapped through a linear layer.

[0042] In wind power output prediction, the input modal units are high-frequency modal units obtained by CEEMD-VMD decomposition and SE classification. The encoder processes each modal unit to extract features that reflect the changing trends and random fluctuations of wind power data. The decoder generates the first output response of each modal unit. Finally, the output responses of the low-frequency modal units are summed to reconstruct the total wind turbine output.

[0043] For low-frequency modal unit modeling, this study considers that the low-frequency modal units obtained after SE classification have strong local stationarity and weak frequency abrupt change characteristics. Therefore, this invention selects GRU for predictive modeling of this type of modal unit. Let the time step be... The input is The output is Hidden state is The GRU mainly consists of two gating units: a reset gate and an update gate, and a candidate state generation module. Its calculation formula is as follows: (twenty three) (twenty four) (25) (26) in, and These are the reset door and the update door, respectively. for The activation function is tanh, which is the hyperbolic tangent function. In the candidate hidden state, , , and , , For trainable parameters, This indicates element-wise multiplication. Update gate. Control the degree to which the previous hidden state is retained, and reset the door. This structure determines the impact of the current input on the hidden state update, enabling it to quickly capture local trend changes in a sequence and avoid overfitting of deep networks to simple dynamic structures. During the training phase, this invention employs a sliding window strategy to construct samples, allowing the model to focus on effectively learning short-term dependencies.

[0044] S6 integrates and reconstructs the wind power prediction results based on the output responses of the first and second modal units.

[0045] Specifically, in this step, the output responses of the first and second modal units are denormalized, and the predicted data after denormalization are summed to obtain and output the wind power prediction results.

[0046] The wind power prediction method based on frequency characteristics provided in this invention proposes a hybrid model that integrates CEEMDAN decomposition, VMD denoising, SE evaluation mechanism, and Transformer-GRU dual-structure modeling. This model effectively reduces high-frequency noise interference when dealing with non-stationary wind turbine output sequences. Furthermore, it fully utilizes frequency characteristics, employing a differentiated learning mechanism design based on the frequency characteristics and complexity differences of the decomposed modal units to improve feature extraction, output accuracy, and time-series dynamic capture capabilities. This enables differentiated modeling and efficient prediction of modal units at different frequencies, effectively improving the accuracy and stability of ultra-short-term wind power prediction, and is suitable for wind power grid-connected scheduling and system operation optimization needs under complex meteorological conditions.

[0047] Figure 2 This is a control structure block diagram of a wind power prediction device provided in an embodiment of the present invention. Figure 2As shown, the wind power prediction device has the function of implementing the aforementioned frequency-oriented wind power prediction method. The device includes a preprocessing module 1, a preliminary decomposition module 2, a secondary decomposition module 3, an evaluation module 4, a modeling and prediction module 5, and a result integration module 6. The preprocessing module 1 is used to obtain the original wind power output sequence of the wind turbine. The preliminary decomposition module 2 is used to decompose the original wind power sequence using the CEEMDAN (Adaptive Noise Complete Set Empirical Mode Decomposition) method to obtain multiple intrinsic mode functions (EMFs). The secondary decomposition module 3 is used to perform secondary decomposition of the multiple EEMFs using the Variational Mode Decomposition (VMD) algorithm to decompose them into multiple modal units with finite bandwidth. The evaluation module 4 is used to evaluate and classify the complexity of the multiple modal units using the Sample Entropy Complexity Evaluation (SE) method to classify the multiple modal units into two categories: high-frequency modal units and low-frequency modal units. Modeling and prediction module 5 is used to model and predict high-frequency modal units using the Transformer model and output the output response of the first modal unit; it uses the GRU model to predict and model low-frequency modal units and output the output response of the second modal unit. Result integration module 6 is used to integrate and reconstruct the wind power prediction results based on the output responses of the first and second modal units.

[0048] The wind power prediction device provided in this invention employs a frequency-characteristic-oriented wind power prediction method. Specifically, it proposes a hybrid model integrating CEEMDAN decomposition, VMD denoising, SE evaluation mechanism, and Transformer-GRU dual-structure modeling. This effectively reduces high-frequency noise interference when processing non-stationary wind turbine output sequences. Furthermore, it fully utilizes frequency characteristics, employing a differentiated learning mechanism design based on the frequency characteristics and complexity differences of the decomposed modal units to improve feature extraction, output accuracy, and time-series dynamic capture capabilities. This enables differentiated modeling and efficient prediction of different frequency modal units, effectively improving the accuracy and stability of ultra-short-term wind power prediction, and is suitable for wind power grid-connected scheduling and system operation optimization needs under complex meteorological conditions.

[0049] Figure 3 This is a schematic diagram of the control structure of another wind power prediction device provided in an embodiment of the present invention. Figure 3 As shown, the control device also includes a processor and a memory for storing processor-executable instructions. The device 4100 can be a computer device, and it may include one or more of the following components: a processor 4101, a memory 4102, a communication interface 4103, and a bus 4104.

[0050] The processor 4101 includes one or more processing cores. The processor 4101 executes various functional applications and information processing by running software programs and modules. The memory 4102 and the communication interface 4103 are connected to the processor 4101 via a bus 4104. The memory 4102 can be used to store at least one instruction, which the processor 4101 uses to execute to implement the various steps in the above method embodiments.

[0051] Furthermore, memory 4102 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, including but not limited to: magnetic disks or optical disks, electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), static random access memory (SRAM), read-only memory (ROM), magnetic storage, flash memory, and programmable read-only memory (PROM).

[0052] For example, in this embodiment of the invention, a non-transitory storage medium including instructions is also provided, such as a memory including instructions, which can be executed by a processor to perform the aforementioned wind power prediction method based on frequency characteristics. For example, the non-transitory storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0053] Figure 5 This is a schematic diagram of wind energy data for La Haute Borne, France in 2017, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of high-frequency modal units after SE classification of modal units provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of low-frequency modal units after SE classification of modal units provided in an embodiment of the present invention; Figure 8 This is a comparison chart of the wind power prediction curve and the actual wind power curve provided in the embodiments of the present invention. For example, refer to... Figures 5 to 8 The technical effects of the aforementioned method will be further described in detail and evidenced below through practical application examples: In this embodiment, an open dataset from the La Haute Borne wind farm in the Grand Est region of northeastern France is used. This wind farm has an installed capacity of 2050 kW, a nacelle height of 80 meters, an altitude of 411 meters, and a rotor diameter of 82 meters. Wind power data from January 1, 2017 to January 1, 2018 are selected, with a sampling interval of 10 minutes, totaling 52,560 data points. Since wind speed and wind power density vary significantly across different months of the year, this invention divides the data by month, constructs corresponding datasets for each month, and performs decomposition-integration prediction on each monthly dataset. The prediction task uses data from the previous hour to predict the wind power output for the next 10 minutes, with the training and test sets divided in a 4:1 ratio. (Reference) Figure 5 The data shown is wind energy data from French wind farms in 2017.

[0054] To verify the effectiveness of the proposed method, all experiments in this invention were conducted on a PC configured with an Intel(R) Core(TM) i9-14900HX@2.20 GHz and 16.00GB RAM, using Python 3.11. Before predicting wind power output, the hyperparameters of the CEEMDAN-VMD-SE-Transformer-GRU model need to be configured. For ease of comparison, Table 1 shows the specific parameter initialization settings for the GRU and Transformer models. The model input is the wind power output data from the past hour, predicting the wind power output 10 minutes later. Dimensionality is adjusted through the last fully connected layer. The Adam optimizer is used during training, and the loss function is the mean squared error (MSE). To ensure the effectiveness of the model, the network architecture remains consistent regardless of whether CEEMDAN-VMD decomposition is introduced.

[0055] Table 1 To measure the predictive performance of the model, this invention employs four evaluation metrics: mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 These four evaluation indicators are defined as follows: (32) (33) (34) (35) in, Indicates the output response. Represents the actual value. This represents the average of the true values. Since the data contains zero values, it is used... As a variant of MAPE.

[0056] CEEMDAN decomposes raw wind power data into several intrinsic mode functions (IMFs) and a fluctuation residual function to reduce data nonlinearity and thus improve prediction accuracy. Taking February power data from French wind farms as an example, analysis of the decomposed mode units revealed that IMF0 exhibits highly irregular and chaotic characteristics, indicating that it contains significant high-frequency noise and short-term fluctuations.

[0057] In the initial prediction of each component of the CEEMDAN decomposition, it was further observed that the prediction accuracy of IMF0 was significantly lower than that of other components. To alleviate this problem, this invention applies VMD denoising processing to IMF0 to form a new denoised_vmd sequence, which is then integrated into other modal units obtained from the CEEMDAN decomposition (such as IMF2 to the fluctuation residual term) to form a new set of modal units. The denoised_vmd sequence after VMD processing and other IMF sequences from the CEEMDAN-VMD decomposition are shown in the February French wind farm data.

[0058] Table 2 To further evaluate and classify the complexity of modal elements, this study uses SE (Sequential Eigenfunction) to quantify all modal elements. Taking February wind power data as an example, the SE values ​​of each modal element are calculated, and the results are shown in Table 2 above. Based on the SE values, modal elements can be divided into high-frequency intrinsic mode functions and low-frequency intrinsic mode functions, as shown in Table 2 above. Figure 6 combine Figure 7 As shown, this is to facilitate subsequent model training.

[0059] To verify the effectiveness of the proposed Transformer-GRU model under CEEMD-SE data preprocessing, nine benchmark methods were selected for comparison. These include Support Vector Regression (SVR), Random Forest (RF), Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU), Transformer, Adaptive Noise Complete Ensemble Empirical Mode Decomposition-Gated Recurrent Unit (CEEMDAN-GRU), CEEMDAN-Transformer, CEEMDAN and Variational Mode Decomposition (VMD)-Gated Recurrent Unit, CEEMDAN-VMD-Transformer, and the proposed CEEMDAN-VMD-SE-Transformer-GRU model. The first five are single models, while the latter five are hybrid prediction models combining data decomposition and integration. Each of these ten methods is denoted as M1-M10. The comparison results on the French wind farm dataset are shown in the table below.

[0060] Table 3 Based on the proposed algorithm, predictions were made for February data, and the results are as follows: Figure 8 As shown.

[0061] Based on the analysis of the experimental results in Table 3, the following conclusions can be drawn: a. In single-model comparisons, GRU and Transformer models outperform SVR, RF, and LSTM in all evaluation metrics (MAPE, RMSE, MAE, R²), mainly due to their stronger ability to capture temporal dynamic structures and express nonlinear features.

[0062] b. Wind power output data based on CEEMDAN decomposition significantly outperforms the unprocessed raw data in prediction performance. According to the data in the table, the annual average MAPE of the CEEMDAN-based GRU and Transformer models is 1.82% and 2.05%, respectively, a reduction of 49.72% and 44.29% compared to the single GRU and Transformer models. This method effectively improves the quality of the input data by decomposing the high volatility and non-stationarity in the original sequence into multiple relatively stationary modal units, thereby enhancing the modeling accuracy of each modal unit and significantly improving the overall prediction performance in the integrated prediction results stage.

[0063] c. Compared to single-pass CEEMD decomposition, the quadratic decomposition of CEEMD with VMD exhibits superior prediction performance. According to the results in the table, the average annual MAPE of GRU and Transformer under the quadratic decomposition model is reduced by 21.43% and 19.02%, respectively, compared to the single-pass decomposition model. This result demonstrates that VMD has significant advantages in noise reduction and improved decomposition accuracy.

[0064] d. The CEEMD-VMD-SE-Transformer-GRU model outperforms all benchmark methods. Among all benchmark methods, the CEEMD-VMD-SE-Transformer-GRU model proposed in this invention achieves the best performance on multiple evaluation metrics such as MAPE, RMSE, MAE, and R², demonstrating excellent output accuracy. This model fully integrates the decomposition advantages of CEEMD in handling nonlinear and non-stationary signals, the effective denoising capability of VMD, the complexity measurement function of SE, and the complementary advantages of Transformer and GRU in local feature extraction and long-term time series dependency modeling, thereby constructing a wind power prediction framework with high accuracy and high robustness.

[0065] Table 4 To further verify the effectiveness and advancement of the proposed CEEMD-VMD-SE-Transformer-GRU model, two representative hybrid prediction models were selected for comparison: the CNN-LSTM model integrating adaptive noise complete set empirical mode decomposition (CEEMDAN-CNN-LSTM) and the EEMD-GRU model combining empirical mode decomposition and gated recurrent units. These two models are highly representative in the field of wind power prediction, effectively capturing the nonlinear dynamic characteristics of time series and are widely used in prediction tasks in wind power and other new energy fields. The comparative experimental results are shown in Table 4. The proposed model outperforms the other two methods in all evaluation metrics, demonstrating superior prediction performance. Compared to CEEMD-CNN-LSTM, the annual average values ​​of MAPE, RMSE, and MAE are reduced by 43.13%, 41.73%, and 42.96%, respectively, while compared to EEMD-GRU, they are reduced by 24.53%, 19.14%, and 24.17%. The improvement in model performance is mainly due to the fusion of multiple methods: CEEMDAN provides an effective decomposition of nonlinear and non-stationary original wind power signals, VMD further suppresses noise interference during the decomposition process, SE introduces a measure of modal unit complexity and classifies modal units; Transformer's advantage in capturing local features of high-frequency modal units and GRU's advantage in modeling the long-term dependence of low-frequency modal units enable more accurate fitting of the dynamic pattern of wind power changes during time series data modeling.

[0066] The above results further demonstrate that the proposed CEEMDAN-VMD-SE-Transformer-GRU model exhibits good generalization ability and superior prediction performance when dealing with wind power data, showing its practical application value in ultra-short-term wind power forecasting.

[0067] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The terms “connected” or “linked” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0068] The above description is merely an optional embodiment of the present invention and is not intended to limit the present invention. 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 power prediction method based on frequency characteristics, characterized in that, include: Step 1: Obtain the original wind power output sequence of the wind turbine. Step 2: The original wind power sequence is decomposed using the adaptive noise complete set empirical mode decomposition method to obtain multiple intrinsic mode functions; Step 3: Use the variational mode decomposition algorithm to perform a secondary decomposition on the multiple intrinsic mode functions to decompose them into multiple modal units with finite bandwidth. Step 4: Use the sample entropy complexity evaluation method to evaluate and classify the complexity of the multiple modal units, so as to divide the multiple modal units into two categories: high-frequency modal units and low-frequency modal units; Step 5: Use the Transformer model to model and predict the high-frequency modal units, and output the first modal unit output response; use the GRU model to predict and model the low-frequency modal units, and output the second modal unit output response. Step six: Based on the output responses of the first and second modal units, integrate and reconstruct the final wind power prediction results.

2. The wind power prediction method based on frequency characteristics according to claim 1, characterized in that, Step one includes: The original wind turbine output sequence corresponding to the historical sampling time of the wind farm is obtained. Missing values ​​are processed, outliers are removed and normalized to obtain the original wind power output sequence.

3. The wind power prediction method based on frequency characteristics according to claim 2, characterized in that, Step two includes: To the original wind power sequence Add to White noise with a mean of 0 ,generate A new wind power sequence ; The empirical mode decomposition method was used to analyze each group of wind power sequences. Decompose to obtain Each component This leads to the first eigenmode function. and abnormal signal components ; To the layer abnormal signal component Add adaptive noise Furthermore, the empirical mode decomposition method is used to further decompose and mean-value the data to obtain the second intrinsic mode function. and new residual fluctuation signals ; Recursion to find the first Individual eigenmode functions and the A residual signal of fluctuation ; Repeat the above steps until the first... A residual signal of fluctuation If the function is monotonic, the adaptive noise complete set empirical mode decomposition method terminates.

4. The wind power prediction method based on frequency characteristics according to claim 2, characterized in that, Step four includes: In a given embedding dimension and similar threshold Under the condition of reconstructing the embedding space of the modal unit, the statistical dimension is then calculated. and The logarithm of the matching vector is used to calculate the logarithmic ratio of the multiple modal units. The average value of the logarithmic ratios corresponding to the multiple modal units is set as the evaluation threshold. Modal units with logarithmic ratios higher than the evaluation threshold are classified as high-frequency modal units, and modal units with logarithmic ratios lower than the evaluation threshold are classified as low-frequency modal units.

5. The wind power prediction method based on frequency characteristics according to claim 2, characterized in that, Step five includes: The encoder of the Transformer model is used to convert the high-frequency modal units into high-dimensional feature representations to capture short-term fluctuations and long-term dependencies. The decoder of the Transformer model generates the first modal unit output response corresponding to each of the high-frequency modal units based on the short-term fluctuations and the long-term dependencies.

6. The wind power prediction method based on frequency characteristics according to claim 2, characterized in that, Step six includes: The output responses of the first and second modal units are denormalized, and the predicted data after denormalization are summed to obtain and output the wind power prediction result.

7. The wind power prediction method based on frequency characteristics according to claim 2, characterized in that, The original wind turbine output sequence is the actual wind power data of the wind turbines in the wind farm.

8. A wind power prediction device, characterized in that, The device includes: The preprocessing module (1) is used to obtain the original wind power output sequence of the wind turbine. The preliminary decomposition module (2) is used to decompose the original wind power sequence using the adaptive noise complete set empirical mode decomposition method to obtain multiple intrinsic mode functions; The secondary decomposition module (3) is used to perform secondary decomposition on the multiple intrinsic mode functions using the variational mode decomposition algorithm, so as to decompose them into multiple mode units with finite bandwidth. Evaluation module (4) is used to evaluate and classify the complexity of multiple modal units using the sample entropy complexity evaluation method, so as to divide the multiple modal units into two categories: high-frequency modal units and low-frequency modal units; The modeling and prediction module (5) is used to model and predict the high-frequency modal unit using the Transformer model and output the output response of the first modal unit; and to predict and model the low-frequency modal unit using the GRU model and output the output response of the second modal unit. The result integration module (6) is used to integrate and reconstruct the final wind power prediction result based on the output response of the first modal unit and the output response of the second modal unit.

9. A wind power prediction device, characterized in that, The device includes: processor; Memory used to store instructions that can be executed by the processor; The processor is configured to execute the frequency-oriented wind power prediction method according to any one of claims 1 to 7.

10. A storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, they implement the wind power prediction method based on frequency characteristics as described in any one of claims 1 to 7.

Images