A two-stage multi-step wind power prediction method driven by meteorological fuzzy factors

Abstract

This invention relates to a two-stage multi-step wind power prediction method driven by meteorological fuzzy factors. The method includes: preprocessing measured wind farm data to obtain a dataset; constructing power prediction samples based on the dataset and performing variational mode decomposition on the standardized power sequences in the samples to obtain intrinsic mode functions (IMFs); predicting each IMF through a gated loop to obtain the first-stage power prediction result, and constructing an error sequence and error prediction samples based on this result and the actual data; decomposing the error prediction samples using CEEMDAN and introducing a multi-level wavelet channel cross-attention mechanism to model the total error prediction sequence; and dynamically weighting and correcting the total prediction error using a fuzzy correction mechanism combined with meteorological factors to obtain the annual prediction sequence. This method maintains relatively consistent prediction performance under different time scales and seasonal conditions, significantly improving prediction reliability and interpretability.

Description

Technical Field

[0001] This invention relates to a wind power prediction method, specifically a two-stage multi-step wind power prediction method driven by meteorological fuzzy factors, belonging to the field of wind power prediction technology. Background Technology

[0002] With the transformation of the energy structure, wind power, as a clean and renewable energy source, is playing an increasingly important role in the power system. However, wind power output is affected by various factors such as meteorological environment and topography, exhibiting strong nonlinear, non-stationary, and random fluctuation characteristics. This uncertainty not only increases the difficulty of grid dispatch and operation but also places higher demands on the high proportion of wind power grid connection and absorption capacity. Therefore, improving the accuracy and stability of wind power forecasting has become an important research direction and engineering challenge in the field of new energy.

[0003] Currently, wind power prediction methods mainly include physical methods, statistical methods, machine learning methods, and deep learning methods. Physical methods are typically based on atmospheric physics laws, using meteorological data and topographic parameters to simulate wind field distribution and wind turbine response. However, they are highly dependent on input data, and the models are complex and computationally expensive. Statistical methods build prediction models by mining the linear correlations of historical data; however, these methods are generally based on linear assumptions and struggle to accurately characterize the complex nonlinear relationship between wind speed and power.

[0004] With the development of machine learning technology, algorithms such as Support Vector Machine (SVM), Decision Tree, Random Forest, and Artificial Neural Network (ANN) have been widely applied in the field of wind power forecasting. Related research institutions have proposed multi-input short-term load forecasting models based on the Random Forest algorithm, and further proposed an improved short-term load forecasting model (RF-DL-RST) combining rough set theory. These methods alleviate the difficulties of nonlinear modeling to some extent, but still suffer from problems such as high dependence on sample quality and insufficient generalization ability.

[0005] In recent years, deep learning models have become a research hotspot due to their powerful feature extraction and time series modeling capabilities. Recurrent Neural Networks (RNNs) and their variants, such as Long Short-Term Memory Networks (LSTMs) and Gated Recurrent Units (GRUs), have shown significant advantages in processing time series data. For example, genetic algorithms can be used to optimize LSTM network structures to achieve high-precision short-term predictions; or the accuracy of wind speed prediction can be improved through the Gray Wolf Optimization (GWO) algorithm and data compression methods; furthermore, K-means++, snake optimization algorithms, and GRUs can be combined for photovoltaic power prediction under different weather conditions. These studies demonstrate that deep learning can effectively capture the complex dynamic characteristics of wind speed and power, but it still faces problems such as high parameter sensitivity and insufficient interpretability.

[0006] Meanwhile, the combination of signal processing and decomposition / reconstruction techniques (such as VMD, EMD, and CEEMDAN) with deep learning has further enhanced the model's ability to model non-stationary signals. Among these, the short-term wind energy prediction model based on a combination of Variational Mode Decomposition (VMD) and ConvLSTM shows significant effectiveness in capturing spatiotemporal features; combining Empirical Mode Decomposition (EMD) with Echo State Networks (ESN) also outperforms single decomposition methods. Furthermore, utilizing CEEMDAN and attention mechanisms for feature extraction and noise suppression can improve the robustness of prediction results. However, decomposition-prediction-reconstruction models are typically complex in structure and computationally intensive, and their prediction accuracy remains insufficient when facing abrupt changes such as extreme weather.

[0007] Furthermore, introducing Physical Information Neural Networks (PINNs) and their extended structures into existing prediction models can enhance the physical interpretability of the models. Specifically, the introduction of Multi-Kernel Pseudo-Inverse Neural Networks (MKPINNs) can improve prediction efficiency and accuracy; the combination of physical constraints and deep learning frameworks can achieve high-precision modeling of wind fields and photovoltaic temperatures. These methods perform well in terms of mechanism fusion and generalization, but still suffer from problems such as complex model design and difficulties in parameter tuning.

[0008] In summary, among existing technologies, hybrid prediction methods are clearly an important research direction for improving wind power prediction performance. However, they still suffer from problems such as insufficient accuracy, complex models, and difficulty in optimization. Static compensation for system errors alone is often insufficient to cope with the volatility under non-stationary operating conditions and extreme weather conditions. This not only leads to error accumulation and prediction bias but also fails to achieve technical objectives such as simplified models and continuous optimization. Summary of the Invention

[0009] To address the problems of existing technologies, this invention provides a two-stage multi-step wind power prediction method driven by meteorological fuzzy factors. This method employs a two-stage framework of "master prediction-error compensation," using multi-level wavelet channel cross-attention (WCXA) to enhance multi-scale feature extraction and meteorological fuzzy factor adaptive correction (FC) to achieve dynamic error compensation, ultimately completing multi-scale wind power prediction. Furthermore, this method mainly focuses on the WCXA attention mechanism and the FC fuzzy correction module, which synergistically solve the shortcomings of existing technologies in feature extraction and error correction.

[0010] To achieve the above technical objectives, this invention provides a two-stage multi-step wind power prediction method based on meteorological fuzzy factors;

[0011] Step S1: Collect measured data from the wind farm, preprocess it, and obtain the dataset;

[0012] Step S2: Construct power prediction samples based on the dataset, and perform variational mode decomposition on the standardized power sequences in the samples to obtain eigenmode functions with finite bandwidth.

[0013] Step S3: Predict each intrinsic mode function through a gated cyclic unit to obtain the power prediction results of the first stage, and construct an error sequence and error prediction samples based on the prediction results and the real data.

[0014] Step S4: Decompose the error prediction samples using CEEMDAN, perform encoder-decoder GRU time-series modeling with multi-level wavelet channel cross-attention, and obtain the total error prediction sequence.

[0015] Step S5: Through the fuzzy correction mechanism, the total forecast error is dynamically weighted and corrected in combination with meteorological factors to obtain the full-year forecast sequence.

[0016] The method provided by this invention mainly divides the prediction process into two stages. The first stage is the process of establishing the master prediction model, which adopts "VMD decomposition + GRU structure" to extract the main trend of wind power time series. The second stage is the establishment of an error compensation model. Based on the master prediction model, "wavelet multi-scale decomposition + channel cross-attention" is used to deeply model the error series in the master prediction model, so that high-frequency disturbances and low-frequency structures in the error can be fully extracted and utilized. By introducing a meteorological fuzzy factor adaptive correction mechanism, the prediction error is dynamically weighted and corrected, thereby ensuring that the method maintains relatively consistent prediction performance under different time scales and different seasonal conditions, and significantly improving the prediction reliability and interpretability.

[0017] As a preferred embodiment, the preprocessing procedure is as follows:

[0018] Step S1-1: Use the average value of adjacent time points to analyze the measured data of the wind farm regarding... The missing power at a given moment is interpolated, and the calculation process is as follows:

[0019] Formula 1: ;

[0020] Step S1-2: Divide the measured data of the wind farm into four parts according to the season: spring, summer, autumn and winter. Then divide the data of each part into training set 1, training set 2 and test set according to the ratio of 6~8:1~3:1.

[0021] In Equation 1, It is a moment The wind power output value; It is a moment The wind power output value; It is a moment The wind power output value.

[0022] It is important to note that wind power exhibits significant seasonal variations. Meteorological conditions, such as temperature, air pressure, wind speed, and humidity, differ considerably across seasons, leading to significant differences in the operating status and output characteristics of wind turbines. If the data from the entire year is directly mixed for modeling, the model will struggle to accurately capture the patterns within each season, resulting in decreased prediction accuracy. Therefore, this invention divides and models the data according to the seasons, enabling the model to adaptively learn for the meteorological characteristics of different seasons, thereby improving overall prediction performance. Spring is from March to May, summer is from June to August, autumn is from September to November, and winter is from December to February.

[0023] As a preferred embodiment, the power prediction sample construction process is as follows: Supervised learning samples are generated using a sliding window. The input features only contain historical standardized power values ​​within the sliding window, and the output is the standardized power values ​​for multiple consecutive steps after the window. Assuming the window length is S and the prediction step size is H, the sample format is:

[0024] Formula 2: ;

[0025] Formula 3: ;

[0026] The process of performing variational mode decomposition on the standardized power sequence in the sample is as follows: separating the multi-scale features of the power to obtain K intrinsic modes (IMFs), and the decomposition objective function and constraints are:

[0027] Formula 4: ;

[0028] In equations 2-4: The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size. Standardized power sequence, For the first One IMF, For its center frequency, This represents the convolution operation; K is the number of modes in the VMD decomposition. Let be the Dirac distribution function.

[0029] Wind power signals exhibit significant nonlinear and nonstationary characteristics, influenced by factors such as sudden wind speed changes, meteorological conditions, and terrain disturbances. Power sequences often simultaneously contain low-frequency trend terms, high-frequency random fluctuations, and mid-frequency transition features. Directly modeling the original power sequence makes it difficult for the model to effectively distinguish the dynamic patterns of different frequency components, leading to decreased prediction accuracy and stability. Therefore, this invention introduces Variational Mode Decomposition (VMD) to adaptively decompose the complex wind power signal into several intrinsic mode functions (IMFs) with finite bandwidth. Each mode corresponds to the characteristic changes of the signal at different time scales. After VMD decomposition, the model can model and predict different mode features separately, and then reconstruct the overall prediction result, thereby significantly improving the modeling accuracy and generalization ability of nonstationary sequences.

[0030] As a preferred approach, the process for obtaining the power prediction results in the first stage is as follows: For each IMF, a GRU predictor is constructed to learn and predict the results of the next H steps, and trained using power prediction samples. The steps are as follows:

[0031] Step S3-1: The GRU predictor learns the historical dependencies of the IMF through the update gate and the reset gate, and extracts time-series features. The process is as follows:

[0032] Formula 5: ;

[0033] Formula 6: ;

[0034] Formula 7: ;

[0035] Formula 8: ;

[0036] Step S3-2: Map the GRU hidden states to H-step predicted values ​​using a fully connected layer, simplifying the structure while maintaining the accuracy of the main trend. The process is as follows:

[0037] Formula 9: ;

[0038] Step S3-3: Linearly superimpose all IMF prediction sequences, and then perform inverse normalization to obtain the power prediction result. The process is as follows:

[0039] Formula 10: ;

[0040] In equations 5-10, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. It is a trainable weight matrix that is input into the gating matrix. For bias terms, For the GRU hidden layer dimension, for Activation function This is element-wise multiplication; For the first A predicted sequence of IMFs, Indicates a fully connected layer; This is the power prediction result.

[0041] As a preferred embodiment, the process for constructing the error sequence and error prediction samples is as follows:

[0042] Through real power With power prediction results The error sequence is obtained. The process of constructing error prediction samples is as follows:

[0043] Formula 11: ;

[0044] Formula 12: ;

[0045] Formula 13: ;

[0046] In equations 11-13, This represents the actual power. For power prediction results, It is an error sequence; The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size.

[0047] As a preferred embodiment, the process of time-series modeling of the encoder-decoder GRU with multi-level wavelet channel cross-attention is as follows: Step S4-1, error prediction samples Error IMFs components were obtained through Ceemdan decomposition. ;

[0048] Step S4-2: Perform discrete wavelet transform on the error sequence to obtain time-frequency features of different frequency bands, and map them to a unified latent space representation to provide multi-layer feature channels for the cross-attention mechanism.

[0049] Step S4-3: Perform multi-level channel cross-attention modeling, assign weights across error IMF channels and wavelet scales, and further decode after weighted fusion to obtain the final prediction error. Then reconstruct the total error prediction sequence.

[0050] As a preferred embodiment, the process of assigning weights to the cross-error IMF channel and wavelet scale is as follows:

[0051] Formula 14: ;

[0052] Formula 15: ;

[0053] Formula 16: ;

[0054] Formula 17: ;

[0055] Formula 18: ;

[0056] Formula 19: ;

[0057] Formula 20: ;

[0058] The weighted fusion process is as follows:

[0059] Equation 21: ;

[0060] Equation 22: ;

[0061] The reconstruction process of the total error prediction sequence is as follows:

[0062] Equation 23: ;

[0063] In equations 14-22, Q represents the coding feature. For time step, For feature dimension, For a moment Feature representation; The wavelet decomposition level is denoted as . For the first Each wavelet channel at time Feature representation; These are the projection matrices for the query, key, and value, respectively. The projected query vector, For the first Key and value representation of layer wavelet channels; The feature dimensions are divided into The dimension of attention head; For the first The input of attention head; The attention weight matrix characterizes the correlation between time steps; It is a scaling factor used to stabilize the gradient; For the first The height in Output on each channel; g i For channel output The gating factor For learnable parameters, For bias terms; for The final prediction error is obtained by further processing the data through a decoder. This is the total error prediction sequence. Further optimization, to reduce computational overhead, sets parameter sharing, i.e., all wavelet layers share the same parameters. .

[0064] After initial prediction, the error sequence between the model output and the true value usually still contains nonlinear and non-stationary features not captured by the main model. These residual errors often exhibit complex multi-scale characteristics, making it difficult to achieve ideal results by directly performing secondary modeling. Virtual Dynamics (VMD) performs excellently in modeling the original wind power sequence, achieving stable frequency domain separation and clear mode division, making it suitable for extracting main trends and energy features. However, the error sequence usually has small amplitude, violent fluctuations, and strong randomness, exhibiting obvious nonlinear and non-stationary characteristics. In this case, VMD is sensitive to noise and struggles to fully decompose complex high-frequency disturbances. In contrast, CEEMDAN, by introducing an adaptive white noise ensemble averaging strategy, can effectively suppress mode aliasing while preserving signal details, improving the stability and completeness of the decomposition. It is more suitable for analyzing error signals dominated by noise and with irregular spectral distribution. Therefore, this invention uses CEEMDAN to finely decompose the residuals in the error correction stage, achieving complementarity between trend modeling and stochastic correction, thereby improving the overall prediction accuracy and robustness.

[0065] As a preferred option, the process of dynamically weighting and correcting the total prediction error by combining meteorological factors is as follows: Step S5-1, by integrating the combined effects of wind shear and wind direction diffusion among meteorological factors through a two-dimensional rule matrix, the dynamic correction coefficient is mapped out.

[0066] Step S5-2: Based on the dynamic correction coefficient, generate a gated correction factor by weighted averaging, so that the compensation range can be adaptively adjusted according to the intensity of the working condition.

[0067] As a preferred embodiment, the process for obtaining the dynamic correction coefficient is as follows:

[0068] Formula 24: ;

[0069] Formula 25: ;

[0070] Equation 26: ;

[0071] Equation 27: ;

[0072] Equation 28: ;

[0073] Equation 29: ;

[0074] Formula 30: ;

[0075] The process of adaptively adjusting the compensation range according to the intensity of the working condition is as follows:

[0076] Equation 31: ;

[0077] Equation 32: ;

[0078] Equation 33: ;

[0079] In equations 24 to 33, For wind speed shear, To spread according to wind direction, and Indicates height , Wind speed at the location; among which and Indicates height , The wind direction at that location Represents modulo operation; and These represent the horizontal mappings of wind speed shear and wind direction diffusion between 0 and 1; Based on the two-dimensional rule matrix, , It is the left boundary. It is a vertex with a membership degree of 1. It is the right boundary; The combined weights under each fuzzy rule, As a weighting adjustment factor, The corrected power prediction results; This represents the wind speed shear at point i. Indicate the wind direction diffusion at point i; For fuzzy membership functions; Let l represent the fuzzy membership degree of the wind speed shear at point l; Let m be the fuzzy membership degree of the wind speed shear at point m; Let h be the fuzzy membership degree of the wind speed shear at point h; This is the total error prediction sequence.

[0080] In wind power forecasting, model errors are often not independent random disturbances, but rather exhibit significant operating condition dependence and seasonality. When wind shear increases or wind direction diffusion intensifies, flow field stability decreases and turbulence intensifies, leading to changes in the residual structure and amplitude. Traditional linear correction or hard thresholding methods are difficult to adapt to different operating conditions and may even introduce abrupt changes and boundary sensitivity issues.

[0081] To address this, this invention employs a fuzzy rule-based approach for adaptive error correction. It represents "low-medium-high" operating conditions with continuous weights through membership functions, ensuring a smooth transition in correction as operating conditions change. A two-dimensional rule matrix is ​​used to synthesize the combined effects of wind shear and wind direction diffusion, mapping out dynamic correction coefficients. Finally, a weighted average is used to generate a gated correction factor, allowing the compensation magnitude to adaptively adjust with the intensity of the operating condition. Compared to black-box models, fuzzy rules offer a clear structure, interpretable parameters, ease of incorporating domain knowledge, and effectively reduce the risk of overfitting.

[0082] As a preferred embodiment, the process of obtaining the full-year forecast sequence is as follows: the power forecast results obtained in spring, summer, autumn, and winter are concatenated to obtain the sequence, and the formula is:

[0083] Equation 34: ;

[0084] in, For the full year forecast series, For spring power forecast results, The power forecast results for summer. For the autumn power forecast results, This is the power forecast result for winter.

[0085] Compared with the prior art, the beneficial technical effects of the technical solution provided by the present invention are as follows:

[0086] 1) The method provided by this invention achieves a comprehensive improvement in prediction stability, model efficiency and engineering deployability through a two-stage collaborative framework consisting of "lightweight master forecast + multi-scale error compensation + meteorological fuzz correction". This structure can maintain relatively consistent prediction performance under different time scales and different seasonal conditions, greatly improve prediction reliability and interpretability, and has significant engineering applicability and wide applicability.

[0087] 2) In the technical solution provided by the present invention, the “VMD decomposition + GRU structure” is adopted in the first stage and used as the main prediction model. It has the advantages of being lightweight, efficient and suitable for real-time deployment on site. Among them, since VMD can effectively separate the main trend components of the power sequence and GRU can stably model the time series relationship, the combination of the two can accurately capture the overall trend of wind power change with low model complexity, providing a stable and reliable basic prediction result for subsequent error compensation.

[0088] 3) In the technical solution provided by this invention, in the second stage, the error sequence in the main prediction model is deeply modeled by “wavelet multi-scale decomposition + channel cross attention”, so that high-frequency disturbances and low-frequency structures in the error can be fully extracted and utilized; among them, the cross attention mechanism can automatically identify the importance of different error channels, dynamically allocate feature weights, and realize the emphasis on key error information, thereby improving the fine-grained characterization of error compensation, so that the final prediction result can maintain better stability under the condition of drastic fluctuations or complex weather changes.

[0089] 4) The technical solution provided by this invention introduces a meteorological fuzzy correction mechanism based on wind shear and wind direction diffusion. Through continuous membership functions and rule reasoning, the error compensation weight is automatically adjusted, so that the correction process can change smoothly with meteorological conditions. This avoids the over- or under-correction caused by the traditional fixed coefficient method when the conditions change abruptly. In addition, since the fuzzy rules are based on the physical mechanism of the wind field, this correction method has good interpretability, which is conducive to engineers' understanding and optimization, and improves the adaptability and robustness of the model under complex meteorological conditions. Attached Figure Description

[0090] Figure 1 This is a schematic diagram of the overall flow of the prediction method provided in Embodiments 1-3 of the present invention;

[0091] Figure 2 This is a schematic diagram of the sliding window principle in the prediction methods provided in Embodiments 1-3 of the present invention;

[0092] Figure 3 This is a schematic diagram of the GRU model structure in the prediction methods provided in Embodiments 1-3 of the present invention;

[0093] Figure 4 The flowchart of CEEMDAN in the prediction method provided in Embodiments 1-3 of this invention is shown below;

[0094] Figure 5 This is a schematic diagram of the multi-level channel cross-attention mechanism in the prediction methods provided in Embodiments 1-3 of the present invention;

[0095] Where n is the input sequence Length, For the input sequence Length, The dimension of the input feature. For the dimension of the query vector, Let be the dimension of the key vector. The dimension of the value vector. To represent the weight matrix, The projection weight of the query. The projection weights of the Key (the first one) (size) Projection weights for Value (the first) (size) For querying the matrix, The key matrix (the first one) (size) Value matrix (the first) (size) Let Q be the dot product of Q and K, and Softmax be the normalization function. Detailed Implementation

[0096] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them.

[0097] The implementation environment of the embodiments provided by this invention is as follows:

[0098] Hardware: Intel Core i5-13400KF processor, 32GB DDR4 memory, NVIDIA GTX 5060 graphics card (8GB VRAM);

[0099] Software: Python 3.9, PyTorch 2.7.1, PyWavelets 1.3.0.

[0100] Embodiments 1-3 of this invention all employ a two-stage multi-step wind power prediction method driven by meteorological fuzzy factors, specifically including the following process:

[0101] Step S1: Collect measured data from the wind farm, preprocess it, and obtain the dataset;

[0102] Step S2: Construct power prediction samples based on the dataset, and perform variational mode decomposition on the standardized power sequences in the samples to obtain eigenmode functions with finite bandwidth.

[0103] Step S3: Predict each intrinsic mode function through a gated cyclic unit to obtain the power prediction results of the first stage, and construct an error sequence and error prediction samples based on the prediction results and the real data.

[0104] Step S4: Decompose the error prediction samples using CEEMDAN, perform encoder-decoder GRU time-series modeling with multi-level wavelet channel cross-attention, and obtain the total error prediction sequence.

[0105] Step S5: Through the fuzzy correction mechanism, the total forecast error is dynamically weighted and corrected in combination with meteorological factors to obtain the full-year forecast sequence.

[0106] As a preferred option, the measured data from wind farms includes wind power series and meteorological factor series; among which, the wind power series... Sampling interval 15 min, N is the total number of samples; meteorological factor sequence: wind speed at 10m altitude 50m height wind speed 10m height wind direction 50m height wind direction .

[0107] As a preferred embodiment, the preprocessing procedure is as follows:

[0108] Step S1-1: Use the average value of adjacent time points to analyze the measured data of the wind farm regarding... The missing power at a given moment is interpolated, and the calculation process is as follows:

[0109] Formula 1: ;

[0110] Step S1-2: Divide the measured data of the wind farm into four parts according to the season: spring, summer, autumn and winter. Then divide the data of each part into training set 1, training set 2 and test set according to 7:2:1.

[0111] In Equation 1, It is a moment The wind power output value; It is a moment The wind power output value; It is a moment The wind power values ​​are shown. Training set 1 is used for the first stage of modeling, and training set 2 is used for the second stage of modeling.

[0112] As a preferred embodiment, the power prediction sample construction process is as follows: Supervised learning samples are generated using a sliding window. The input features only contain historical standardized power values ​​within the sliding window, and the output is the standardized power values ​​for multiple consecutive steps after the window. Assuming the window length is S and the prediction step size is H, the sample format is:

[0113] Formula 2: ;

[0114] Formula 3: ;

[0115] The process of performing variational mode decomposition on the standardized power sequence in the sample is as follows: separating the multi-scale features of the power to obtain K intrinsic modes (IMFs), and the decomposition objective function and constraints are:

[0116] Formula 4: ;

[0117] In equations 2-4: The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size. Standardized power sequence, For the first One IMF, For its center frequency, This represents the convolution operation; K is the number of modes in the VMD decomposition. Let be the Dirac distribution function.

[0118] As a preferred approach, the process for obtaining the power prediction results in the first stage is as follows: For each IMF, a GRU predictor is constructed to learn and predict the results of the next H steps, and trained using power prediction samples. The steps are as follows:

[0119] Step S3-1: The GRU predictor learns the historical dependencies of the IMF through the update gate and the reset gate, and extracts time-series features. The process is as follows:

[0120] Formula 5: ;

[0121] Formula 6: ;

[0122] Formula 7: ;

[0123] Formula 8: ;

[0124] Step S3-2: Map the GRU hidden states to H-step predicted values ​​using a fully connected layer, simplifying the structure while maintaining the accuracy of the main trend. The process is as follows:

[0125] Formula 9: ;

[0126] Step S3-3: Linearly superimpose all IMF prediction sequences, and then perform inverse normalization to obtain the power prediction result. The process is as follows:

[0127] Formula 10: ;

[0128] In equations 5-10, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. It is a trainable weight matrix that is input into the gating matrix. For bias terms, For the GRU hidden layer dimension, for Activation function This is element-wise multiplication; For the first A predicted sequence of IMFs, Indicates a fully connected layer; This is the power prediction result.

[0129] As a preferred embodiment, the process for constructing the error sequence and error prediction samples is as follows:

[0130] Through real power With power prediction results The error sequence is obtained. The process of constructing error prediction samples is as follows:

[0131] Formula 11: ;

[0132] Formula 12: ;

[0133] Formula 13: ;

[0134] In equations 11-13, This represents the actual power. For power prediction results, It is an error sequence; The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size.

[0135] As a preferred embodiment, the process of time-series modeling of the encoder-decoder GRU with multi-level wavelet channel cross-attention is as follows: Step S4-1, error prediction samples Error IMFs components were obtained through Ceemdan decomposition. ;

[0136] Step S4-2: Perform discrete wavelet transform on the error sequence to obtain time-frequency features of different frequency bands, and map them to a unified latent space representation to provide multi-layer feature channels for the cross-attention mechanism.

[0137] Step S4-3: Perform multi-level channel cross-attention modeling, assign weights across error IMF channels and wavelet scales, and further decode after weighted fusion to obtain the final prediction error. Then reconstruct the total error prediction sequence.

[0138] As a preferred embodiment, the process of assigning weights to the cross-error IMF channel and wavelet scale is as follows:

[0139] Formula 14: ;

[0140] Formula 15: ;

[0141] Formula 16: ;

[0142] Formula 17: ;

[0143] Formula 18: ;

[0144] Formula 19: ;

[0145] Formula 20: ;

[0146] The weighted fusion process is as follows:

[0147] Equation 21: ;

[0148] Equation 22: ;

[0149] The reconstruction process of the total error prediction sequence is as follows:

[0150] Equation 23: ;

[0151] In equations 14-22, Q represents the coding feature. For time step, For feature dimension, For a moment Feature representation; The wavelet decomposition level is denoted as . For the first Each wavelet channel at time Feature representation; These are the projection matrices for the query, key, and value, respectively. The projected query vector, For the first Key and value representation of layer wavelet channels; The feature dimensions are divided into The dimension of attention head; For the first The input of attention head; The attention weight matrix characterizes the correlation between time steps; It is a scaling factor used to stabilize the gradient; For the first The height in Output on each channel; g i For channel output The gating factor For learnable parameters, For bias terms; for The final prediction error is obtained by further processing the data through a decoder. This is the total error prediction sequence.

[0152] As a preferred option, the process of dynamically weighting and correcting the total prediction error by combining meteorological factors is as follows: Step S5-1, by integrating the combined effects of wind shear and wind direction diffusion among meteorological factors through a two-dimensional rule matrix, the dynamic correction coefficient is mapped out.

[0153] Step S5-2: Based on the dynamic correction coefficient, generate a gated correction factor by weighted averaging, so that the compensation range can be adaptively adjusted according to the intensity of the working condition.

[0154] As a preferred embodiment, the process for obtaining the dynamic correction coefficient is as follows:

[0155] 1. Meteorological factors: wind speed shear and wind direction diffusion The calculation is as follows:

[0156] Formula 24: ;

[0157] Formula 25: ;

[0158] 2. Using functions Map the shear and diffusion levels to a range of 0-1.

[0159] Equation 26: ;

[0160] 3. Multiply the shear level and wind direction diffusion level by the defined modified intensity base matrix. This results in... Basic two-dimensional rule matrix :

[0161] Equation 27: ;

[0162] 4. Based on quantiles (Q10 / Q50 / Q90) and trigonometric membership functions (Equation 22), wind shear and wind direction diffusion are classified into low / medium / high levels:

[0163] Equation 28: ;

[0164] For wind shear, the classification is defined as follows:

[0165] Equation 29: ;

[0166] For wind direction diffusion, the division is defined as follows:

[0167] Formula 30: ;

[0168] The process of adaptively adjusting the compensation range according to the intensity of the working condition is as follows:

[0169] Equation 31: ;

[0170] Equation 32: ;

[0171] Equation 33: ;

[0172] In equations 24 to 33, For wind speed shear, To spread according to wind direction, and Indicates height , Wind speed at the location; among which and Indicates height , The wind direction at that location Represents modulo operation; and These represent the horizontal mappings of wind speed shear and wind direction diffusion between 0 and 1; Based on the two-dimensional rule matrix, , It is the left boundary. It is a vertex with a membership degree of 1. It is the right boundary; The combined weights under each fuzzy rule, As a weighting adjustment factor, The corrected power prediction results; This represents the wind speed shear at point i. Indicate the wind direction diffusion at point i; For fuzzy membership functions; Let l represent the fuzzy membership degree of the wind speed shear at point l; Let m be the fuzzy membership degree of the wind speed shear at point m; Let h be the fuzzy membership degree of the wind speed shear at point h; This is the total error prediction sequence.

[0173] As a preferred embodiment, the process of obtaining the full-year forecast sequence is as follows: the power forecast results obtained in spring, summer, autumn, and winter are concatenated to obtain the sequence, and the formula is:

[0174] Equation 34: ;

[0175] in, For the full year forecast series, For spring power forecast results, The power forecast results for summer. For the autumn power forecast results, This is the power forecast result for winter.

[0176] Example 1

[0177] The dataset consists of 70,176 samples, with a time scale of 15 minutes, using power data from a 99MW wind farm in Inner Mongolia from 2019 to 2020. Wind speed and direction data were also collected simultaneously.

[0178] The parameters were set as follows: MD parameter K=5, learning rate 0.001, batch size 32, number of training epochs 500; WCXA, wavelet scale 3.

[0179] Training was performed using the above prediction method. During sample generation, S=24 and H=1. The prediction results after 15 minutes (1-step) showed that the prediction method had NRMSE=1.69 and NMAE=0.98, which were 6.31% and 4.20% lower than the existing GRU method, respectively.

[0180] Example 2

[0181] This embodiment is exactly the same as Embodiment 1, except that S=24 and H=4 in the sample generation process. The prediction results after 1 hour (4-step) show that NRMSE=2.66 and NMAE=1.65, which are reduced by 9.11% and 6.71% respectively compared with the GRU method in the prior art.

[0182] Example 3

[0183] This embodiment is exactly the same as Embodiment 1, except that S=48 and H=16 in the sample generation process. In the 4h (16-step) prediction results, NRMSE=5.10 and NMAE=3.27, which are 10.10% and 8.75% lower than the GRU method in the prior art, respectively.

[0184] The present invention compares the effects of Examples 1-3 with those of common prediction methods in the prior art, and the results are shown in Table 1.

[0185] ;

[0186] Table 1 shows the two-stage predictions. The first stage uses GRU, and the table indicates the method used in the second stage. CA represents channel attention, XA represents cross-attention, WCXA represents multi-level wavelet channel cross-attention, and FC represents fuzzy adaptive correction. The prediction results in Table 1 are from ablation experiments. For different steps of the experiment, the values ​​on the left are the evaluation results, and the values ​​on the right are the percentage improvement compared to GRU. As shown in Table 1, the prediction method provided by this invention has significant advantages over existing prediction methods. Under the same computing power, the evaluation result improvement rate can reach up to 10.1%.

Claims

1. A two-stage multi-step wind power prediction method based on meteorological fuzzy factors, characterized in that: Step S1: Collect measured data from the wind farm, preprocess it, and obtain the dataset; Step S2: Construct power prediction samples based on the dataset, and perform variational mode decomposition on the standardized power sequences in the samples to obtain eigenmode functions with finite bandwidth. Step S3: Predict each intrinsic mode function through a gated cyclic unit to obtain the power prediction results of the first stage, and construct an error sequence and error prediction samples based on the prediction results and the real data. Step S4: Decompose the error prediction samples using CEEMDAN, perform encoder-decoder GRU time-series modeling with multi-level wavelet channel cross-attention, and obtain the total error prediction sequence. Step S5: Through the fuzzy correction mechanism, combined with meteorological factors, the total forecast error is dynamically weighted and corrected to obtain the full-year forecast sequence. The process of dynamically weighting and correcting the total prediction error by combining meteorological factors is as follows: Step S5-1: The dynamic correction coefficient is mapped by comprehensively analyzing the combined effects of wind shear and wind direction diffusion among meteorological factors through a two-dimensional rule matrix. Step S5-2: Based on the dynamic correction coefficient, generate a gated correction factor by weighted averaging, so that the compensation range can be adaptively adjusted according to the intensity of the working condition.

2. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 1, characterized in that: The preprocessing process is as follows: Step S1-1: Use the average value of adjacent time points to analyze the measured data of the wind farm regarding... The missing power at a given moment is interpolated, and the calculation process is as follows: Formula 1: ; Step S1-2: Divide the measured data of the wind farm into four parts according to the season: spring, summer, autumn and winter. Then divide the data of each part into training set 1, training set 2 and test set according to (6~8):(1~3):

1. In Equation 1, It is a moment The wind power output value; It is a moment The wind power output value; It is a moment The wind power output value.

3. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 2, characterized in that: The process of constructing the power prediction samples is as follows: Supervised learning samples are generated using a sliding window. The input features only contain the historical standardized power values ​​within the sliding window, and the output is the standardized power values ​​for multiple consecutive steps after the window. Assuming the window length is S and the prediction step size is H, the sample format is: Formula 2: ; Formula 3: ; The process of performing variational mode decomposition on the standardized power sequence in the sample is as follows: separating the multi-scale features of the power to obtain K intrinsic modes (IMFs), and the decomposition objective function and constraints are: Formula 4: ; In equations 2-4: The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size. Standardized power sequence, For the first One IMF, For its center frequency, This represents the convolution operation; K is the number of modes in the VMD decomposition. Let be the Dirac distribution function.

4. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 3, characterized in that: The process of obtaining the power prediction results in the first stage is as follows: For each IMF, a GRU predictor is constructed to learn and predict the results of the next H steps, and is trained using power prediction samples. The steps are as follows: Step S3-1: The GRU predictor learns the historical dependencies of the IMF through the update gate and the reset gate, and extracts time-series features. The process is as follows: Formula 5: ; Formula 6: ; Formula 7: ; Formula 8: ; Step S3-2: Map the GRU hidden states to H-step predicted values ​​using a fully connected layer, simplifying the structure while maintaining the accuracy of the main trend. The process is as follows: Formula 9: ; Step S3-3: Linearly superimpose all IMF prediction sequences, and then perform inverse normalization to obtain the power prediction result. The process is as follows: Formula 10: ; In equations 5-10, To update the door, To reset the door, In the candidate hidden state, This is the final hidden state. It is a trainable weight matrix that is input into the gating matrix. For bias terms, For the GRU hidden layer dimension, for Activation function This is element-wise multiplication; For the first A predicted sequence of IMFs, Indicates a fully connected layer; This is the power prediction result.

5. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 4, characterized in that: The process of constructing the error sequence and error prediction samples is as follows: Through real power With power prediction results The error sequence is obtained. The process of constructing error prediction samples is as follows: Formula 11: ; Formula 12: ; Formula 13: ; In equations 11-13, This represents the actual power. For power prediction results, It is an error sequence; The window for inputting the model has a length of [length missing]. ; The multi-step prediction results output by the model have a length of S is the input window length; H is the prediction step size.

6. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 5, characterized in that: The process of temporal modeling of the encoder-decoder GRU that integrates multi-level wavelet channel cross-attention is as follows: Step S4-1, error prediction samples Error IMFs components were obtained through Ceemdan decomposition. ; Step S4-2: Perform discrete wavelet transform on the error sequence to obtain time-frequency features of different frequency bands, and map them to a unified latent space representation to provide multi-layer feature channels for the cross-attention mechanism. Step S4-3: Perform multi-level channel cross-attention modeling, assign weights across error IMF channels and wavelet scales, and further decode after weighted fusion to obtain the final prediction error. Then reconstruct the total error prediction sequence.

7. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 6, characterized in that: The process of assigning weights to the IMF channel and wavelet scale across the error is as follows: Formula 14: ; Formula 15: ; Formula 16: ; Equation 17: ; Formula 18: ; Formula 19: ; Formula 20: ; The weighted fusion process is as follows: Equation 21: ; Equation 22: ; The reconstruction process of the total error prediction sequence is as follows: Equation 23: ; In equations 14-22, Q represents the coding feature. For time step, For feature dimension, For a moment Feature representation; The wavelet decomposition level is denoted as . For the first Each wavelet channel at time Feature representation; These are the projection matrices for the query, key, and value, respectively. The projected query vector, The key and value representations for the wavelet channels of the nth layer; The feature dimensions are divided into The dimension of attention head; For the first The input of attention head; The attention weight matrix characterizes the correlation between time steps; It is a scaling factor used to stabilize the gradient; For the first The height in Output on each channel; g i For channel output The gating factor For learnable parameters, For bias terms; for The final prediction error is obtained by further processing the data through a decoder. This is the total error prediction sequence.

8. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 1, characterized in that: The process of obtaining the dynamic correction coefficient is as follows: Formula 24: ; Formula 25: ; Equation 26: ; Equation 27: ; Formula 28: ; Equation 29: ; Formula 30: ; The process of adaptively adjusting the compensation range according to the intensity of the working condition is as follows: Equation 31: ; Equation 32: ; Equation 33: ; In equations 24 to 33, For wind speed shear, To spread according to wind direction, and Indicates height , Wind speed at the location; among which and Indicates height , The wind direction at that location Represents modulo operation; and These represent the horizontal mappings of wind speed shear and wind direction diffusion between 0 and 1; Based on the two-dimensional rule matrix, , It is the left boundary. It is a vertex with a membership degree of 1. It is the right boundary; The combined weights under each fuzzy rule, As a weighting adjustment factor, The corrected power prediction results; This represents the wind speed shear at point i. Indicate the wind direction diffusion at point i; For fuzzy membership functions; Let l represent the fuzzy membership degree of the wind speed shear at point l; Let m be the fuzzy membership degree of the wind speed shear at point m; Let h be the fuzzy membership degree of the wind speed shear at point h; This is the total error prediction sequence.

9. The two-stage multi-step wind power prediction method based on meteorological fuzzy factors as described in claim 8, characterized in that: The process of obtaining the full-year forecast sequence is as follows: the power forecast results obtained in spring, summer, autumn and winter are concatenated to obtain the sequence, and the formula is: Equation 34: ; in, For the full year forecast series, For spring power forecast results, The power forecast results for summer. For the autumn power forecast results, This is the power forecast result for winter.

Images