A method for predicting wind power

Abstract

This invention discloses a method for predicting wind power, comprising the following steps: acquiring and preprocessing wind power time series data; extracting local statistical characteristics of the preprocessed data using a sliding time window and constructing a non-stationarity index; adaptively adjusting the number of modes and penalty factor of variational mode decomposition using the non-stationarity index, and decoupling the original wind power time series data into intrinsic mode function components through the variational mode decomposition with adjusted parameters; extracting high-frequency transient fluctuation features of high-frequency mode components using a long short-term memory network; extracting global long-term trend features of low-frequency mode components using a sparse self-attention mechanism; and performing cross-modal cross-attention fusion of the high-frequency transient fluctuation features and the global long-term trend features to obtain the wind power prediction result. This invention provides a more accurate and stable solution for multimodal time series modeling of wind power.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, and more specifically, relates to a method for predicting wind power. Background Technology

[0002] With the rapid development of new energy technologies globally, wind energy, as a highly practical and promising renewable energy source, is an indispensable part of the future energy system. However, wind power generation is easily affected by multiple factors such as terrain, topography, environment, and weather, resulting in wind power sequences characterized by strong non-stationarity and complex fluctuations. Therefore, high-precision wind power forecasting plays a crucial role in ensuring the stable operation of the power system and improving grid acceptance capacity.

[0003] Currently, wind power prediction methods are gradually evolving towards deep learning models, such as Long Short-Term Memory (LSTM) networks and Informer models. However, wind power exhibits complex correlations across multiple time scales, and existing single models often struggle to simultaneously capture both short-term and long-term dependencies, frequently leading to error accumulation due to mode aliasing and insufficient modeling of long-range dependencies. In recent years, Variational Mode Decomposition (VMD) has gained widespread attention due to its excellent signal decoupling capabilities. However, in practical applications, fixing the number of modes (K value) in VMD can easily lead to high-frequency noise residue or loss of low-frequency trends due to improper parameter selection, thus limiting the decomposition effect. Existing hybrid models, such as LSTM-Informer or fixed VMD-Informer combinations, still suffer from inflexible parameter control and local prediction lag when dealing with highly non-stationary abrupt signals.

[0004] A prior art patent (CN117828547A) proposes a method for predicting short-term wind power. The method includes: optimizing the parameters of Variational Mode Decomposition (VMD) using an optimization algorithm; decomposing wind power data using the optimized VMD; constructing a prediction model to predict the decomposed subsequences; and reconstructing the prediction results of each subsequence. This scheme mainly uses an optimization algorithm to globally optimize the VMD parameters, and its parameter determination method focuses more on offline search for the optimal solution, lacking online adaptive adjustment capabilities for the local time-varying characteristics and non-stationary intensity changes of wind power time series. When wind power is under highly non-stationary conditions such as rapid rise, sudden drop, or oscillating switching, the VMD parameters fixed to a single optimization result are difficult to change synchronously with the signal complexity, easily leading to over-decomposition in the stationary phase and under-decomposition in the abrupt phase, thus affecting the physical interpretability of the modal components and the subsequent prediction accuracy. Meanwhile, in the process of predicting the decomposed subsequences, the scheme does not further distinguish the modeling requirements of high-frequency transient fluctuation information and low-frequency long-term trend information, making it difficult to take into account both the ability to capture local mutations and the ability to represent long-range dependencies. This results in problems such as local prediction lag, error accumulation, and insufficient collaborative modeling of features across multiple time scales in complex non-stationary scenarios. Summary of the Invention

[0005] To address the issues of strong non-stationarity and significant local time-varying characteristics in wind power time series, and the lack of flexible adjustment capabilities for the determination of parameters in existing variational mode decomposition methods to accommodate the time-varying characteristics of the signal, which leads to local prediction lag and limited overall prediction accuracy, a wind power prediction method is proposed.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows: This invention provides a method for predicting wind power, comprising the following steps: The wind power time series data were acquired and preprocessed. The local statistical characteristics of the preprocessed data were extracted using a sliding time window and a non-stationarity index was constructed. The non-stationarity index is used to determine the number of modes and the penalty factor of variational mode decomposition, and variational mode decomposition is performed on the original wind power time series data to output intrinsic mode function components, which include high-frequency mode components and low-frequency mode components. High-frequency transient fluctuation features of high-frequency modal components are extracted using long short-term memory networks; global long-term trend features of low-frequency modal components are extracted using sparse self-attention mechanisms. Cross-modal cross-attention fusion is performed on the high-frequency transient fluctuation characteristics and global long-term trend characteristics to obtain wind power prediction results.

[0007] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention innovatively constructs a non-stationarity index, overcoming the limitations of traditional variational mode decomposition (VMD) with fixed modal parameters. Starting from the non-stationarity evolution mechanism, this invention establishes an adaptive VMD parameter optimization strategy, which automatically adjusts the number of modes and penalty factors according to the time-varying characteristics of the signal. This effectively avoids the omission of high-frequency features under complex abrupt changes and the over-decomposition problem in stationary phases, significantly improving the physical interpretability of the decomposition results. Furthermore, this invention achieves collaborative modeling of features across multiple time scales and dual-channel decoupling, deeply integrating the sensitivity of LSTM to transient high-frequency fluctuations and the advantages of Informer's sparse self-attention mechanism for modeling global long-term trends. Supplemented by adaptive weight allocation through cross-modal fusion, it successfully solves the problems of error accumulation and short-term fluctuation lag in long-term prediction using a single model. Experiments show that the method of this invention achieves a high degree of consistency between the predicted trajectory and the actual value in non-stationary scenarios such as rapid power increases and oscillating switching, providing a more accurate and stable solution for multi-modal time-series modeling of wind power. Attached Figure Description

[0008] To make the objectives and technical solutions of this invention clearer, the following drawings are provided and described: Figure 1 This is a flowchart of a wind power prediction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the Long Short-Term Memory (LSTM) network structure used in the embodiments of the present invention; Figure 3 This is a schematic diagram of cross-modal attention fusion used in the embodiments of the present invention.

[0009] Figure 4 This is a graph showing the relationship between the wind power signal and the change in the adaptive mode number K in an embodiment of the present invention; Figure 5 This is a comparison chart of the prediction performance of a single model without variational mode decomposition and a basic fusion model in an embodiment of the present invention; Figure 6 This is a comparison chart of the prediction performance of the model in this embodiment of the invention and that of a traditional VMD single network; Figure 7 This is a violin plot showing the probability density distribution of prediction errors for different models in embodiments of the present invention. Detailed Implementation

[0010] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0011] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0012] Example 1: This invention provides a method for predicting wind power output, such as... Figure 1 The diagram shows a flowchart of a wind power prediction method. The specific steps are as follows: S1: Obtain wind power time series data and preprocess it. Use a sliding time window to extract local statistical characteristics of the preprocessed data and construct a non-stationarity index.

[0013] More specifically, considering that wind farm sensors may experience communication interruptions or equipment malfunctions during actual operation, resulting in missing values ​​or abrupt changes in the collected wind power data, data preprocessing is required for the raw wind power time series data. Preprocessing of the wind power time series data includes the following steps: S1.1.1: For missing values ​​in the wind power time series data, linear interpolation or the mean of previous and subsequent time points is used to fill in the missing values ​​to obtain preliminary corrected data.

[0014] S1.1.2: For abnormal data in the preliminary corrected data that clearly exceeds the physical extreme value, an anomaly detection method is used for detection, and the detected abnormal data is removed and smoothed out to obtain cleaned data.

[0015] More specifically, the anomaly detection method employs the isolated forest algorithm or the 3σ criterion as the core detection mechanism.

[0016] S1.1.3: To eliminate the influence of different units on the convergence speed of deep learning model training, this embodiment normalizes the cleaned data (e.g., Min-Max normalization), mapping all wind power values ​​to [0,1] or […]. Within the interval [1,1].

[0017] After preprocessing, a sliding window mechanism is used to construct time series samples, and the dataset is divided into training and test sets in a 7:3 ratio for subsequent feature learning and performance verification by the deep learning network.

[0018] More specifically, the preprocessed wind power time series data is subjected to local statistical characteristic extraction and a non-stationarity index is constructed, including the following steps: S1.2.1: Obtain wind power time series data ; S1.2.2: Select a sliding window length of... For the first A sliding window The local fluctuation intensity (LFI) is expressed as follows:

[0019] in, Indicates the first window The mean of all power sample points within the window, the local fluctuation intensity is used to measure the intensity of the power signal oscillation around its mean within the current window.

[0020] S1.2.3: To characterize the trend of wind power change on a macroscopic time scale, calculate the... A sliding window The expression for the Trend Mutation Factor (TMF) is as follows:

[0021] in, Indicates the first The trend abrupt change factor is used to quantify the degree of transition of the power reference level between adjacent windows.

[0022] S1.2.4: This invention considers the contribution of fluctuation amplitude and trend abrupt change to the overall nonstationarity of the signal to be equally important. The local fluctuation intensity and trend abrupt change factor are normalized and weighted and summed to obtain the nonstationarity index. The expression is as follows:

[0023] in, and Let represent the maximum and minimum values ​​in the historical sliding window sample set, respectively. and For the preset weighting coefficients, and In this embodiment, the following is set and .

[0024] The nonstationarity index The range of values ​​is The closer the value of this indicator is to 1, the higher the degree of non-stationarity of the wind power signal during that period, and the richer the high-frequency abrupt change components contained in the signal; conversely, it indicates that the signal tends to be stable.

[0025] S2: The number of modes and penalty factor of variational mode decomposition are determined by using the non-stationarity index, and variational mode decomposition is performed on the original wind power time series data to output intrinsic mode function components, including high-frequency mode components and low-frequency mode components. Variational mode decomposition (VMD) is a non-recursive signal processing technique based on a variational optimization framework. Compared with the frequency band aliasing and endpoint effects that are easy to occur in traditional empirical mode decomposition (EMD), VMD can decompose complex waveforms into multiple mode components with specific center frequencies and finite bandwidths, effectively improving the physical interpretability and algorithm stability of the decomposition results.

[0026] More specifically, traditional methods typically employ fixed parameters determined through manual experience when applying VMD. Inappropriate selection of these fixed parameters can easily lead to residual high-frequency noise (under-decomposition) or loss of low-frequency trends (over-decomposition), severely limiting the decomposition effect under complex operating conditions. To overcome the insufficient response of fixed K-value decomposition to the time-varying nature of signals at different operating stages, this invention introduces an adaptive parameter optimization mechanism based on non-stationarity indices on the traditional VMD framework.

[0027] The method of adaptively adjusting the number of modes and penalty factor in variational mode decomposition using a nonstationarity index includes the following steps: S2.1.1: Preset lower bound for the number of modes in the variational mode decomposition algorithm With upper limit And the basic lower limit of the secondary penalty factor. With upper limit .

[0028] S2.1.2: Utilizing the non-stationarity index Calculate the number of adaptive modalities for the current window. The expression is as follows:

[0029] in, Indicates the window number. This represents the rounding function. Using this formula, when wind power enters a stable phase with gradual fluctuations, i.e. When it approaches 0, Automatically converges to lower orders This avoids spurious modes caused by excessive decomposition; when entering a strongly non-stationary mutation phase, i.e. When it approaches 1, Automatically jump to a higher level This is to refine the high-frequency band division.

[0030] S2.1.3: To adapt to high-dimensional frequency band division and maintain orthogonality between modes, the penalty factor needs to increase with the increase of decomposition dimension. This is achieved by utilizing the non-stationarity index of the current sliding window. Calculate the adaptive penalty factor The expression is as follows:

[0031] In a preferred embodiment of the present invention, the above boundary parameters are set as follows, taking into account the actual spectral characteristics of the wind power signal: , , , This adaptive control mechanism completely overcomes the limitation of traditional low-order decomposition in complex abrupt signals, which suffers from severe under-decomposition. It ensures that the signal can provide a clearly structured temporally decoupled component for the subsequent dual-channel prediction network at any time-varying stage.

[0032] More specifically, the original wind power time series data is decoupled into intrinsic mode function components through variational mode decomposition with adjusted parameters, including the following steps: S2.2.1: Convert the original wind power time series data Decomposed into Modal functions The corresponding center frequency is With the objective of minimizing the sum of bandwidths of all modal components, and satisfying the reconstruction constraint that the superposition of all modal components equals the original signal itself, a variational optimization problem is constructed, expressed as follows:

[0033]

[0034] in, Let k be the time variable and k be the modal index. For the partial derivative operator with respect to time, For Dirac functions, For the convolution operation of signals, The frequency domain separation performance of variational mode decomposition is highly dependent on the set number of modes, which is used to shift the spectrum of each mode to the complex exponential term of the corresponding baseband. and secondary penalty factor The selection.

[0035] S2.2.2: To solve the variational optimization problem with equality constraints, the constrained variational optimization problem is subjected to an augmented Lagrangian transformation, introducing a penalty factor. and Lagrange multipliers The augmented Lagrange function is obtained. The augmented Lagrangian function is used to characterize the joint optimization relationship between the bandwidth constraints of each modal component and the signal reconstruction constraints, and its expression is as follows:

[0036] in, This represents the dot product of two vectors. This represents the square of the L2 norm of a vector.

[0037] S2.2.3: The Alternating Direction Method of Multipliers (ADMM) is used to find the extrema of the augmented Lagrangian function. This is achieved by iteratively updating each modal component, modal center frequency, and Lagrangian multiplier in the frequency domain to find the saddle point (i.e., the optimal solution) of the augmented Lagrangian function. The modal functions... The update expression is as follows:

[0038] in, For frequency variables in the frequency domain, Let be the current iteration number, and i represent the index of the modal components other than the k-th mode. Represents raw wind power time series data Fourier transform form, This represents the sum of the Fourier transforms of all modal components except the k-th modal component. Represents Lagrange multipliers The Fourier transform form of .

[0039] Center frequency The update expression is as follows:

[0040] in, For the first The Fourier transform form of the k-th modal component after the next iteration update.

[0041] Lagrange multipliers The update expression is as follows:

[0042] in, Update the step size factor for the preset time derivative. For the first Fourier transform of the Lagrange multipliers in the next iteration Original wind power time series data Fourier transform form, For the first The sum of the Fourier transforms of all k modal components after each iteration update.

[0043] S2.2.4: Update the modal component results obtained in the current iteration. Center frequency update results and the Lagrange multiplier update results Compare the results with the previous iteration and determine whether the preset convergence condition is met based on the error between two adjacent iterations.

[0044] S2.2.5: If the preset convergence condition is met, the modal components output when the convergence condition is met are determined as the intrinsic mode function components corresponding to the original wind power time series data; if the preset convergence condition is not met, the above iterative update process is repeated using the modal components, center frequency and Lagrange multipliers obtained in the current iteration until the preset convergence condition is reached, and multiple intrinsic mode function components are output.

[0045] After adaptive VMD decomposition, the original complex non-stationary power sequence is dynamically decomposed into multiple low-dimensional IMF subsequences, each component corresponding to a specific frequency feature. This decomposition strategy decouples the signal, allowing the deep learning model to focus on feature learning at different time scales, thereby avoiding feature conflicts caused by a single network processing chaotic sequences simultaneously.

[0046] S3: Utilizing a Long Short-Term Memory (LSTM) network to extract high-frequency transient fluctuation features of high-frequency modal components. The LSTM network is a recurrent neural network specifically designed for processing time-series data. It overcomes the gradient vanishing or exploding problems that occur in traditional RNNs when training long sequences, effectively capturing local dynamics and short-term dependencies in time series data, such as... Figure 2 As shown, the specific process is as follows: S3.1: Divide the multiple intrinsic mode function components output by variational mode decomposition according to their center frequencies, determine the intrinsic mode function components corresponding to the high-frequency bands, and construct a high-frequency input sequence from the high-frequency intrinsic mode function components in chronological order; determine the component value of the high-frequency input sequence at the current time as the input at the current time. ; S3.2: The forget gate control information is retained, utilizing the input at the current time. The hidden layer state at the previous time step Calculate the forget gate state The expression is as follows:

[0047] in, For activation function, Here is the forget gate weight matrix. Forget gate bias parameters; S3.3: Input gate control information is added, utilizing the current input time. The hidden layer state at the previous time step Calculate the input gate state The expression is as follows:

[0048] in, For activation function, The input gate weight matrix, For input gate bias parameters; S3.4: Input using the current time The hidden layer state at the previous time step Calculate the candidate cell state and compare the candidate cell state with the input gate state. Perform weighting, and simultaneously set the forget gate state Cell state compared to the previous moment We perform weighted analysis to obtain the cell state at the current moment. The expression is as follows:

[0049] in, The hyperbolic tangent activation function is used. This is the cell state weight matrix. These are cell state bias parameters; S3.5: The output gate controls the final output information, utilizing the input at the current time. The hidden layer state at the previous time step Calculate the output gate state The expression is as follows:

[0050] in, This is the output gate weight matrix. These are the output gate bias parameters; S3.6: Set the output gate state The current cell state Perform joint computation to obtain the hidden layer state at the current time. and the hidden layer state The expression for the high-frequency transient fluctuation characteristic output corresponding to the current moment is as follows:

[0051] S3.7: Output the hidden layer state at the current moment. Output as the hidden layer state output unit for the next time step. Continue inputting into the Long Short-Term Memory network and repeat the above steps until the temporal features of the high-frequency transient fluctuation components at each time step are extracted, and the high-frequency transient fluctuation features corresponding to the intrinsic mode function components are output.

[0052] S4: Extract global long-term trend features of low-frequency modal components using the sparse self-attention mechanism of the Informer model encoder.

[0053] The specific process is as follows: S4.1: Extract the low-frequency mode components that reflect the long-term trend from the multiple intrinsic mode function components output by variational mode decomposition, and arrange them in chronological order to obtain the low-frequency mode component sequence. S4.2: The low-frequency modal component sequence is sequentially input into an Informer encoder network formed by stacking multiple attention layers with the same structure, wherein each attention layer includes a sparse self-attention mechanism; S4.3: Utilize the sparse self-attention mechanism in the current attention layer to perform long-range dependency modeling on the low-frequency modal component sequence features input to the attention layer, determine the dominant temporal features in the low-frequency modal component sequence that contribute significantly to the global long-term trend representation, and output the trend feature representation corresponding to the current attention layer; S4.4: The trend feature representation output by the current attention layer is input into the next attention layer, and the sparse self-attention mechanism in the next attention layer is used to perform a deeper level of global correlation feature extraction on the trend feature representation. S4.5: Repeat the above layer-by-layer input and sparse self-attention feature extraction process until the low-frequency modal component sequence passes through each attention layer in the Informer model encoder in sequence, and output the final encoded features as the global long-term trend features corresponding to the low-frequency modal components.

[0054] S5: Cross-modal cross-attention fusion is performed on the high-frequency transient fluctuation characteristics and global long-term trend characteristics to obtain multi-step wind power prediction results, such as... Figure 3 As shown, the specific process is as follows: S5.1: Utilize global long-term trend features to guide the global prediction direction, perform linear mapping on the global long-term trend features, and generate a fusion query matrix. The expression is as follows:

[0055] in, This is the first cross-modal learnable weight matrix specific to the multimodal fusion layer; S5.2: Using the high-frequency transient fluctuation features as external context constraints to supplement local details, a linear mapping is performed on the high-frequency transient fluctuation features to obtain the fusion bond matrix. and fusion value matrix The expression is as follows:

[0056]

[0057] in, These are the second and third cross-modal learnable weight matrices, respectively, specific to the multimodal fusion layer.

[0058] S5.3: Utilizing probabilistic sparse attention mechanism to... , , Dynamic weight allocation is performed to obtain the comprehensive feature vector after multi-scale feature fusion, as shown in the following expression:

[0059] in, This is a matrix of low-frequency query vectors that, after probabilistic sparse filtering, are retained and tend to dominate future macroeconomic trends. The scaling factor added to prevent gradient vanishing; The physical significance of this step lies in: allowing the low-frequency trend representing the macroscopic world to... Actively search for and dynamically absorb high-frequency transient fluctuations that match the target. and This allows the model to be more sensitive to local power abrupt changes without disrupting long-term time series dependencies.

[0060] The comprehensive feature vector is used to characterize the dynamic absorption of high-frequency transient fluctuation information by low-frequency trends, thereby enhancing the model's ability to perceive local power abrupt changes without destroying long-range time series dependencies. This allows macroscopic low-frequency trends to dynamically absorb matching high-frequency transient fluctuation information without destroying long-range time series dependencies.

[0061] S5.4: Utilizing comprehensive feature vectors A dual-channel reconstruction of "global trend baseline + local feature compensation" is performed, combined with the projection weight matrix of the fully connected layer. Calculate the local feature correction term, and then combine the local feature correction term with the low-frequency trend feature. direct extrapolation reference value at time The results are superimposed to obtain the multi-step prediction results of wind power. The expression is as follows:

[0062] in, To predict the step size, This is the projection weight matrix of the fully connected layer, used to map the high-dimensional fused features to the actual dimensions of wind power. for The low-frequency trend characteristics at any given time. This reconstruction mechanism provides reliable macroscopic trend support for the prediction in the second term on the right side of the equation, and superimposes high-frequency transient correction on the basis of the first term, finally outputting a full-scale wind power co-prediction result with physical interpretability.

[0063] As a preferred embodiment, in order to ensure that the above-mentioned dual-channel feature extraction layer can be effectively trained and converged, the present invention sets specific parameters and loss functions for its network training process.

[0064] The Adam optimization algorithm (Adaptive Moment Estimation) is used to backpropagate and update the weight matrix and bias parameters of the Long Short-Term Memory Network and the Informer model.

[0065] Regarding hyperparameter settings, the initial learning rate was set to 0.001, and a learning rate decay strategy was adopted to avoid getting stuck in local optima in the later stages of training. The batch size was set to 32, the maximum number of epochs was set to 100, and an early stopping mechanism was introduced. Training was automatically terminated when the loss on the validation set no longer decreased for 10 consecutive epochs to prevent the model from overfitting.

[0066] Ultimately, this invention achieves adaptive weight updates for information in each frequency band, breaking through the limitations of traditional fixed-mode combined static fusion, and outputs a full-scale wind power collaborative prediction result with physical interpretability.

[0067] To verify the effectiveness of this invention, data from a large wind power base in northern my country was collected for testing.

[0068] Specifically, the total installed capacity of the base is 1960.91MW. The output voltage of the wind turbine generators in the base is 690V. After being stepped up to 35kV by a box-type transformer, it is connected to the 35kV bus of the substation through the collector line, and finally stepped up to 220kV by the main transformer and connected to the regional power grid.

[0069] like Figure 4As shown in the figure, the wind power signal and the adaptive mode number K change response relationship are shown in the figure. It can be seen from the figure that when the wind power is in different operating stages such as smooth or drastic change, the non-stationarity index (INS) drives the variational mode decomposition mode number K to dynamically adjust. This shows that the adaptive parameter control mechanism of the present invention can accurately sense the time-varying evolution of the signal, thereby improving the high-frequency feature omission (under-decomposition) under complex sudden change conditions and the over-decomposition problem in the stationary stage, and improving the rationality and physical interpretability of the decomposition results.

[0070] like Figure 5 The figure shows a comparison of the prediction performance of a single model without signal decomposition and a basic fusion model. This figure demonstrates that, without variational mode decomposition preprocessing, the prediction results of LSTM, Informer, and their basic fusion models on the original wind power sequence are still insufficient to adequately adapt to the strong non-stationarity of wind power time series. They are prone to phenomena such as amplitude underestimation, phase lag, and insufficient characterization of abrupt changes. This indicates that in complex wind power scenarios, relying solely on deep learning networks is insufficient to fully achieve multi-timescale feature modeling.

[0071] like Figure 6 The figure shows a comparison of the prediction performance of this invention and the traditional single VMD network. This figure reflects the difference in trajectory fitting between the dual-channel prediction architecture of this invention and the traditional single predictors VMD-LSTM and VMD-Informer in locally high-fluctuation regions, under the condition that both use VMD decomposition. The results show that single networks have limitations in their ability to model peak and valley inflection points or capture short-term mutations, while this invention, through dual-channel collaborative modeling, achieves complementary representation of high-frequency transient fluctuations and low-frequency long-term trends, thus obtaining higher-accuracy prediction results in regions of severe fluctuation.

[0072] like Figure 7 The figure shows a violin plot of the probability density distribution of prediction errors for different models. This plot illustrates the distribution characteristics of error divergence, systematic bias, and extreme outliers in the tail for each prediction model. The results show that the error distribution of the present invention exhibits an extremely sharp needle-like structure and converges tightly to around 0MW, eliminating systematic bias and minimizing the risk of extreme prediction failures.

[0073] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting wind power output, characterized in that, Includes the following steps: The wind power time series data is acquired and preprocessed. Local statistical characteristics are extracted from the preprocessed data and a non-stationarity index is constructed. The non-stationarity index is used to adaptively adjust the number of modes and the penalty factor of variational mode decomposition. The original wind power time series data is decoupled into intrinsic mode function components through variational mode decomposition with adjusted parameters. The intrinsic mode function components include high-frequency mode components and low-frequency mode components. High-frequency transient fluctuation characteristics of high-frequency modal components are extracted using long short-term memory networks; The global long-term trend features of low-frequency modal components are extracted using a sparse self-attention mechanism; Cross-modal cross-attention fusion is performed on the high-frequency transient fluctuation characteristics and global long-term trend characteristics to obtain wind power prediction results.

2. The method for predicting wind power according to claim 1, characterized in that, Preprocessing wind power time series data includes the following steps: For missing values ​​in the wind power time series data, linear interpolation or the mean filling method of the preceding and following time points is used to correct them, and preliminary corrected data is obtained. For abnormal data in the preliminary corrected data that clearly exceeds the physical extreme value, an anomaly detection method is used for detection, and the detected abnormal data is removed and smoothed out to obtain cleaned data. The cleaned data is then normalized to obtain preprocessed data.

3. The method for predicting wind power according to claim 2, characterized in that, The anomaly detection method employs the isolated forest algorithm or the 3σ criterion as the core detection mechanism.

4. The method for predicting wind power according to claim 1, characterized in that, The preprocessed wind power time series data is subjected to local statistical characteristic extraction and non-stationarity index construction, including the following steps: Obtain wind power time series data ; Select the sliding window length as For the first A sliding window Local fluctuation intensity The expression is as follows: in, Indicates the first window The mean of all power sample points within the range; Calculate the first A sliding window Trend mutation factor The expression is as follows: in, Indicates the first The power average within the previous adjacent window or historical reference window of each window; The local fluctuation intensity and the trend abrupt change factor are normalized and weighted and summed to obtain the nonstationarity index. The expression is as follows: in, and Let represent the maximum and minimum values ​​in the historical sliding window sample set, respectively. and For the preset weighting coefficients, and .

5. The method for predicting wind power according to claim 4, characterized in that, The method of adaptively adjusting the number of modes and penalty factor in variational mode decomposition using a nonstationarity index includes the following steps: Preset the basic lower limit value of the number of modes in the variational mode decomposition algorithm With upper limit And the basic lower limit of the secondary penalty factor. With upper limit ; Using nonstationarity index Calculate the number of adaptive modalities for the current window. The expression is as follows: in, Indicates the window number. This represents the rounding function; Using the nonstationarity index of the current sliding window Calculate the adaptive penalty factor The expression is as follows: 。 6. The method for predicting wind power according to claim 5, characterized in that, The original wind power time series data is decoupled into intrinsic mode function components by variational mode decomposition with adjusted parameters, including the following steps: Original wind power time series data Decomposed into Modal functions The corresponding center frequency is With the objective of minimizing the sum of bandwidths of all modal components, and satisfying the reconstruction constraint that the superposition of all modal components equals the original signal itself, a variational optimization problem is constructed, expressed as follows: in, Let k be the time variable and k be the modal index. For the partial derivative operator with respect to time, For the Dirac function, For the convolution operation of signals, This is a complex exponential term used to shift the spectrum of each mode to the corresponding baseband; The variational optimization problem is subjected to an augmented Lagrangian transformation, which introduces a penalty factor. and Lagrange multipliers The augmented Lagrange function is obtained. The expression is as follows: in, This represents the dot product of two vectors. The square of the second norm of a vector; The extrema of the augmented Lagrange function are solved by iteratively updating each modal component, the center frequency of each mode, and the Lagrange multipliers in the frequency domain to find the saddle point of the augmented Lagrange function, where the modal function... The update expression is as follows: in, For frequency variables in the frequency domain, Let be the current iteration number, and i represent the index of the modal components other than the k-th mode. Represents raw wind power time series data Fourier transform form, This represents the sum of the Fourier transforms of all modal components except the k-th modal component. Represents Lagrange multipliers The Fourier transform form; Center frequency The update expression is as follows: in, For the first The Fourier transform form of the k-th modal component after the next iteration update; Lagrange multipliers The update expression is as follows: in, Update the step size factor for the preset time derivative. For the first Fourier transform of the Lagrange multipliers in the next iteration Original wind power time series data Fourier transform form, For the first The sum of the Fourier transforms of all k modal components after each iteration update; Update the modal component results obtained in the current iteration. Center frequency update results and the Lagrange multiplier update results Compare the results with the previous iteration and determine whether the preset convergence condition is met based on the error between two adjacent iterations. If the preset convergence condition is met, the modal components output when the convergence condition is met are determined as the intrinsic mode function components corresponding to the original wind power time series data; if the preset convergence condition is not met, the above iterative update process is repeated using the modal components, center frequency and Lagrange multipliers obtained in the current iteration until the preset convergence condition is reached, and multiple intrinsic mode function components are output.

7. The method for predicting wind power according to claim 6, characterized in that, The method for finding the extrema of the augmented Lagrange function is the alternating direction multiplier method.

8. The method for predicting wind power according to claim 6, characterized in that, High-frequency transient fluctuation characteristics of intrinsic mode function components are extracted using long short-term memory networks, including the following steps: Multiple intrinsic mode function (IMF) components are divided according to their center frequencies. IMF components corresponding to high-frequency bands are identified, and these high-frequency IMF components are constructed into a high-frequency input sequence in chronological order. The component value of this high-frequency input sequence at the current time is determined as the input value at that time. ; Input using the current time The hidden layer state at the previous time step Calculate the forget gate state The expression is as follows: in, For activation function, Here is the forget gate weight matrix. Forget gate bias parameters; Input using the current time The hidden layer state at the previous time step Calculate the input gate state The expression is as follows: in, For activation function, The input gate weight matrix, For input gate bias parameters; Input using the current time The hidden layer state at the previous time step Calculate the candidate cell state and compare the candidate cell state with the input gate state. Perform weighting, and simultaneously set the forget gate state Cell state compared to the previous moment We perform weighted analysis to obtain the cell state at the current moment. The expression is as follows: in, The hyperbolic tangent activation function is used. This is the cell state weight matrix. These are cell state bias parameters; Input using the current time The hidden layer state at the previous time step Calculate the output gate state The expression is as follows: in, This is the output gate weight matrix. These are the output gate bias parameters; The output gate state The current cell state Perform joint computation to obtain the hidden layer state at the current time. and the hidden layer state The expression for the high-frequency transient fluctuation characteristic output corresponding to the current moment is as follows: Output the hidden layer state at the current moment. Output as the hidden layer state output unit for the next time step. Continue inputting into the Long Short-Term Memory network and repeat the above steps until the temporal features of the high-frequency transient fluctuation components at each time step are extracted, and the high-frequency transient fluctuation features corresponding to the intrinsic mode function components are output.

9. The method for predicting wind power according to claim 1, characterized in that, The global long-term trend features of low-frequency modal components are extracted using a sparse self-attention mechanism, including the following steps: The low-frequency mode components that reflect long-term trends are extracted from the multiple intrinsic mode function components output by variational mode decomposition and arranged in chronological order to obtain the low-frequency mode component sequence. The low-frequency modal component sequence is sequentially input into an Informer encoder network formed by stacking multiple attention layers with the same structure, wherein each attention layer includes a sparse self-attention mechanism. By utilizing the sparse self-attention mechanism in the current attention layer, long-range dependency modeling is performed on the low-frequency modal component sequence features input to the attention layer, the dominant temporal features that contribute significantly to the global long-term trend representation in the low-frequency modal component sequence are determined, and the trend feature representation corresponding to the current attention layer is output. The trend feature representation output by the current attention layer is input into the next attention layer, and the sparse self-attention mechanism in the next attention layer is used to perform a deeper level of global correlation feature extraction on the trend feature representation. Repeat the above layer-by-layer input and sparse self-attention feature extraction process until the low-frequency modal component sequence passes through each attention layer in the Informer model encoder in sequence, and output the final encoded features as the global long-term trend features corresponding to the low-frequency modal components.

10. The method for predicting wind power according to claim 1, characterized in that, Cross-modal attention fusion of high-frequency transient fluctuation characteristics and global long-term trend characteristics includes the following steps: A linear mapping is performed on the global long-term trend features to generate a fusion query matrix. The expression is as follows: in, This is the first learnable weight matrix; A linear mapping is performed on the high-frequency transient fluctuation characteristics to obtain the fusion bond matrix. and fusion value matrix The expression is as follows: in, These are the second and third learnable weight matrices, respectively. Using probabilistic sparse attention mechanism to , , Dynamic weight allocation is performed to obtain the comprehensive feature vector after multi-scale feature fusion, as shown in the following expression: in, This is the low-frequency query vector matrix retained after probabilistic sparse filtering. The scaling factor added to prevent gradient vanishing; Using comprehensive feature vectors Perform dual-channel reconstruction, combined with the projection weight matrix of the fully connected layer. Calculate the local feature correction term, and then combine the local feature correction term with the low-frequency trend feature. direct extrapolation reference value at time The results are superimposed to obtain the multi-step prediction results of wind power. The expression is as follows: in, To predict the step size, Here is the projection weight matrix of the fully connected layer. for Low-frequency trend characteristics of time.

Images