Wind power plant power prediction method and system

Abstract

The invention discloses a wind power plant power prediction method and a wind power plant power prediction system, which are used for solving the technical problem of low power prediction precision caused by the conditions of modal aliasing and incomplete decomposition of weak non-stationary components in the existing wind power plant power prediction method. The method comprises the steps of obtaining and preprocessing historical power signals of a wind power plant to obtain target signals; an improved self-adaptive noise complete set empirical mode decomposition method is adopted, and corresponding residual terms and a plurality of intrinsic mode function components are generated in combination with a plurality of white noises; reconstructing a power prediction sequence based on a preset embedding dimension, the residual term and the intrinsic mode function component; and determining an input space sub-region according to the input variable characteristics, selecting a target power prediction model in the sub-region, and outputting a power prediction value of the wind power plant after inputting a sequence.

Description

Technical Field

[0001] This invention relates to the field of power prediction technology, and in particular to a power prediction method and system for wind power plants. Background Technology

[0002] With the increasing demand for refined management and optimized operation of power systems, accurate forecasting of wind power plant power has become crucial. Accurate wind power forecasting helps optimize grid dispatch, improve power supply reliability, and reduce operating costs.

[0003] However, wind power data itself exhibits typical non-stationary and non-linear characteristics, and its variations are influenced by a variety of complex factors such as weather, season, workday type, and residents' lifestyles, making high-precision power forecasting extremely challenging. Currently, most wind power forecasting methods focus on short-term (e.g., daily, weekly) forecasts. While these methods have achieved some success, they still cannot meet the engineering application requirements for grid-connected wind power generation for monthly or even longer-term forecasts.

[0004] Existing wind power prediction methods that employ mainstream machine learning models such as support vector regression can, under normal meteorological conditions and stable wind turbine operation scenarios, basically fit the nonlinear relationship of power signals and provide prediction results with some reference value, thus barely meeting the basic engineering application requirements for power prediction in grid-connected wind power generation. However, these methods lack a dynamic adaptation mechanism specifically designed for non-stationary signals and have a weak ability to capture the switching of statistical characteristics of power signals over time. Even if some methods introduce mode decomposition techniques (such as Empirical Mode Decomposition) to attempt to process non-stationary signals, there are still issues such as mode aliasing and incomplete decomposition of weak non-stationary components, resulting in low power prediction accuracy. Summary of the Invention

[0005] This invention provides a method and system for predicting the power of wind power plants, which solves the technical problems of low power prediction accuracy caused by mode aliasing and incomplete decomposition of weak non-stationary components in existing wind power plant power prediction methods.

[0006] The first aspect of this invention provides a method for predicting the power output of a wind power plant, comprising:

[0007] The historical power signal of the wind power plant is acquired, and the historical power signal of the wind power plant is preprocessed to output the historical power signal of the target wind power plant.

[0008] An improved adaptive noise complete set empirical mode decomposition method is adopted to generate residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises.

[0009] Based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant, the power prediction sequence is reconstructed.

[0010] Based on the input variable characteristics of the historical power signal of the wind power plant, the input space sub-region to which the historical power signal of the wind power plant belongs is determined;

[0011] Select a target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the wind power plant power prediction value.

[0012] Optionally, the preprocessing of the historical power signal of the wind power plant to output the historical power signal of the target wind power plant includes:

[0013] The historical power signals of the wind power plants are processed to remove outliers and complete the data, and the intermediate historical power signals of the wind power plants are output.

[0014] The historical power signals of the intermediate wind power plants are normalized, and the power signals of the target wind power plants are output.

[0015] Optionally, the improved adaptive noise complete set empirical mode decomposition method generates residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises, including:

[0016] Extract the intrinsic mode function component of each of the white noises, and add the intrinsic mode function components of each of the white noises to the historical power signal of the target wind power plant, and output multiple historical power signals of the wind power plant with added components;

[0017] Calculate the local mean of the historical power signal of each of the added component wind power plants, and perform mean operation on the local mean of the historical power signal of each of the added component wind power plants to output the average local mean.

[0018] The intrinsic mode function components of the target wind power plant's power signal are output by subtracting the historical power signal of the target wind power plant from the average local mean.

[0019] Determine whether the local mean becomes a trend term;

[0020] If so, the average local mean corresponding to the trend term will be used as the residual term;

[0021] If not, extract new intrinsic mode function components from each of the white noises and use the average local mean as the new historical power signal of the target wind farm.

[0022] The process jumps to the step of adding the intrinsic mode function components of each of the white noises to the historical power signal of the target wind power plant, and outputting multiple historical power signals of the wind power plant with added components, until the average local mean becomes the trend term.

[0023] Optionally, the step of reconstructing the power prediction sequence based on a preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant includes:

[0024] Based on the preset embedding dimension, multiple embedding vectors are constructed for each intrinsic mode function component corresponding to the historical power signal of the target wind power plant.

[0025] Calculate the distance similarity between each embedding vector of each intrinsic mode function component;

[0026] The number of embedding vectors corresponding to distance similarities less than a preset similarity threshold is counted to obtain the local similarity counts corresponding to each embedding vector of each intrinsic mode function component.

[0027] The local count of each intrinsic mode function component is calculated based on the local similarity count corresponding to each embedding vector of each intrinsic mode function component.

[0028] The preset embedding dimension is updated to determine a new preset embedding dimension;

[0029] Based on the new preset embedding dimension, a new local count is determined for each of the intrinsic mode function components;

[0030] Calculate the sample entropy of each intrinsic mode function component based on the new local count and local count of each intrinsic mode function component;

[0031] The sample entropy of each intrinsic mode function component is compared with a preset sample entropy threshold.

[0032] The intrinsic mode function component corresponding to any sample entropy less than or equal to the preset sample entropy threshold is taken as the target intrinsic mode function component.

[0033] The power prediction sequence is reconstructed based on the target intrinsic mode function components and the residual terms.

[0034] Optionally, the formula for calculating the sample entropy is as follows:

[0035] ;

[0036] in, The sample entropy; For the intrinsic mode function components in the preset embedding dimension, The local count at that time is the new local count of the intrinsic mode function components; For the intrinsic mode function components in the preset embedding dimension, Local counts at time, i.e., local counts of eigenmode function components; Preset embedding dimension; This refers to the updated preset embedding dimension, i.e., the new preset embedding dimension. The length of the time series; This is the similarity tolerance threshold.

[0037] Optionally, the power prediction sequence is specifically:

[0038] ;

[0039] in, For power prediction sequences; The total number of the target intrinsic mode function components; For the k-th target intrinsic mode function component; This is the residual term.

[0040] A second aspect of the present invention provides a power prediction system for wind power plants, comprising:

[0041] The acquisition module is used to acquire historical power signals of wind power plants, preprocess the historical power signals of wind power plants, and output the historical power signals of target wind power plants.

[0042] The generation module is used to generate residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises using an improved adaptive noise complete set empirical mode decomposition method.

[0043] The reconstruction module is used to reconstruct the power prediction sequence based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant.

[0044] The determination module is used to determine the input spatial sub-region to which the historical power signal of the wind power plant belongs based on the input variable characteristics of the historical power signal of the wind power plant.

[0045] The prediction module is used to select a target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the wind power plant power prediction value.

[0046] A computer device provided in a third aspect of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the wind power plant power prediction method as described in any of the preceding claims.

[0047] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the wind power plant power prediction method as described in any of the preceding claims.

[0048] The fifth aspect of the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the wind power plant power prediction method as described in any of the preceding claims.

[0049] As can be seen from the above technical solutions, the present invention has the following advantages:

[0050] The above-mentioned technical solution of the present invention provides a method for predicting the power of a wind power plant. The method acquires historical power signals from the wind power plant, preprocesses these signals, and outputs the historical power signal of the target wind power plant. An improved adaptive noise complete set empirical mode decomposition method is used to generate residual terms and multiple intrinsic mode function (IMF) components corresponding to the historical power signal of the target wind power plant based on the historical power signal and multiple white noises. A power prediction sequence is reconstructed based on a preset embedding dimension and the residual terms and IMF components corresponding to the historical power signal of the target wind power plant. Finally, the historical power signal of the wind power plant is determined based on the input variable characteristics of the historical power signal. The input space sub-region is defined; a target power prediction model is selected within the input space sub-region, and the power prediction sequence is input into the target power prediction model for prediction, outputting the wind power plant power prediction value; based on the above scheme, this invention optimizes the noise injection strategy and mode screening mechanism through an improved decomposition method, which can effectively suppress the mode aliasing phenomenon commonly found in traditional mode decomposition, while enhancing the thoroughness of decomposition of weak non-stationary components, so that the reconstructed power prediction sequence completely retains the dynamic law and detailed features of the original signal; combined with the input space sub-region division and target model selection based on historical power features, the prediction model can be specifically adapted to power sequences with different laws, thereby improving the power prediction accuracy. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the steps of a wind power plant power prediction method provided in Embodiment 1 of the present invention;

[0053] Figure 2 This is an overall framework diagram of a wind power plant power prediction method provided in Embodiment 1 of the present invention;

[0054] Figure 3 This is a structural block diagram of a wind power plant power prediction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0055] This invention provides a method and system for predicting the power of wind power plants, which addresses the technical problems of low power prediction accuracy caused by mode aliasing and incomplete decomposition of weak non-stationary components in existing wind power plant power prediction methods.

[0056] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0057] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a wind power plant power prediction method provided in Embodiment 1 of the present invention.

[0058] This invention provides a method for predicting the power output of a wind power plant, comprising:

[0059] Step 101: Obtain the historical power signal of the wind power plant, preprocess the historical power signal of the wind power plant, and output the historical power signal of the target wind power plant.

[0060] It should be noted that the acquired "historical power signal of wind power plant" refers to the time-series data recorded by monitoring equipment during the past operating cycle of wind power plant, which reflects the change of power generation over time. It is the original information carrier for subsequent analysis. Preprocessing it yields a historical power time-series sequence (historical power signal of target wind power plant) with data quality that meets the analysis requirements, providing a regular and reliable input for the subsequent mode decomposition process.

[0061] Specifically, step 101 may include the following sub-steps:

[0062] S11. Remove outliers and complete data in the historical power signal of the wind power plant, and output the intermediate historical power signal of the wind power plant.

[0063] S12. Normalize the historical power signal of the intermediate wind power plant and output the power signal of the target wind power plant.

[0064] It should be noted that, firstly, outlier detection should be performed on the historical power signals of wind power plants. This can be done using a box plot method based on interquartile range to identify outliers that significantly deviate from the main data distribution, or by applying a moving window standard deviation method. This involves first calculating the statistical characteristics of the data within a local window (calculating the standard deviation of the data within the local window), and then detecting and removing abnormal peaks, troughs, or zero-value drift points that significantly exceed a set standard deviation threshold. After outlier cleaning and removal, data completion is necessary to address missing values ​​in the original data and gaps created after outlier removal. The specific filling method should be selected based on the length of the missing period: linear interpolation should be prioritized for filling short-term, consecutively missing data points; if there are longer missing data segments, or if the power data itself exhibits obvious periodicity, the mean or median power of similar operating days at the same time in the same historical period from the same wind power plant should be used for filling, thus ensuring the integrity of the time series and providing reliable basic data for subsequent analysis.

[0065] Furthermore, this invention uses the Z-SCORE method (Z-score standardization) to convert the preprocessed wind power plant power data (i.e., intermediate wind power plant power signals) into a standard normal distribution with a mean of 0 and a standard deviation of 1. The purpose is to eliminate the dimensional differences in power output between different wind power plants and to address the problem of outliers in meter data collection. This method can be expressed as:

[0066] ;

[0067] in, This represents the average power signal of the intermediate wind power plants; The standard deviation of the power signal from the intermediate wind power plant; The i-th power data value after standardization in the power signal of the target wind power plant; This is the i-th raw power data value in the power signal of the intermediate wind power plant.

[0068] In this embodiment, two methods are used to detect outliers and missing data in the raw power data of wind power plants: First, a box plot method based on interquartile range is used to identify outliers deviating from the main distribution range; second, a standard deviation method based on a sliding window is used to remove outliers such as peaks, valleys, and zero drift by setting a standard deviation multiple threshold. Then, data completion is performed: linear interpolation is used to fill in missing data over short periods; while for longer missing periods or data with obvious periodicity, the mean or median power of the same wind power plant under similar historical operating conditions is used to ensure the continuity of the power time series. To eliminate the influence of differences in units or acquisition accuracy between different wind power plants, the preprocessed data is standardized using the Z-SCORE method, transforming it into a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby improving the robustness and generalization ability of the model training.

[0069] Step 102: Using the improved adaptive noise complete set empirical mode decomposition method, based on the historical power signal of the target wind power plant and multiple white noises, generate the residual term and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant.

[0070] It should be noted that this invention proposes an improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) method to decompose k sub-time series. This method is an integrated empirical decomposition algorithm specifically designed to handle non-stationary signals and nonlinear data, and is used for multi-day wind power forecasting. It is an improvement on EEMD (Ensemble Empirical Mode Decomposition) by introducing adaptive noise and an improved IMF (Intrinsic Mode Function) extraction method. Its aim is to correct the shortcomings of traditional signal decomposition through an adaptive noise injection strategy, providing support for accurate signal separation. This decomposition technique decodes each time series signal into a finite number of IMFs and incorporates white noise patterns into the original signal to enhance the quality of the IMFs, effectively reducing residual noise. ICEEMDAN can handle adaptive noise, optimizing noise injection location and adaptively attenuating noise amplitude; ICEEMDAN can correct frequency aliasing, achieving noise-induced band separation and residual focusing decomposition.

[0071] In this invention, the ICEEMDAN algorithm is used to decompose the power time-series signal into k intrinsic mode functions (IMFs) and residual terms. ICEEMDAN achieves high-frequency structure enhancement and frequency aliasing correction by introducing adaptive amplitude-controlled white noise and weighting the first IMF.

[0072] Specifically, step 102 may include the following sub-steps:

[0073] S21. Extract the intrinsic mode function component of each white noise, and add the intrinsic mode function component of each white noise to the historical power signal of the target wind power plant, and output multiple historical power signals of the wind power plant with added components.

[0074] S22. Calculate the local mean of the historical power signal of each added component wind power plant, and perform mean operation on the local mean of the historical power signal of each added component wind power plant, and output the average local mean.

[0075] S23. Subtract the historical power signal of the target wind power plant from the average local mean, and output the intrinsic mode function components of the power signal of the target wind power plant.

[0076] S24. Determine whether the local mean of the average is a trend term;

[0077] S25. If so, the average local mean corresponding to the trend term will be used as the residual term.

[0078] S26. If not, extract new intrinsic mode function components from each white noise and use the average local mean as the new target wind power plant historical power signal.

[0079] S27. Jump to execute the step of adding the intrinsic mode function components of each white noise to the historical power signal of the target wind power plant and outputting multiple historical power signals of the wind power plant with added components, until the average local mean becomes the trend term.

[0080] The trend term refers to the remaining component in the iterative process of modal decomposition of the historical power signal of the target wind power plant, when the average local mean no longer contains the local fluctuation components that can be decomposed into intrinsic mode functions (IMFs), and only reflects the overall trend of the original power signal (such as a long-term slow rise, fall, or stable overall trend). It is the sign that the decomposition iteration has ended and is ultimately retained as a residual term, representing the overall trend characteristics of the power signal that cannot be further decomposed.

[0081] It should be noted that white noise is introduced to assist in the decomposition of EMD (Empirical Mode Decomposition) to address the mode aliasing problem. However, instead of directly adding the white noise, the first IMF of the white noise is added. This is a key difference between ICEEMD and EEMD. EMD is performed on the white noise itself, extracting its first IMF (i.e., extracting the L-th IMF, where L is the IMF number, L≥1). This IMF has high-frequency characteristics and can mimic the high-frequency structure of the signal without disturbing the low-frequency components. Then, the intrinsic mode function components of each white noise are added to the historical power signal of the target wind farm, outputting multiple historical power signals of the wind farm with added components.

[0082] ;

[0083] in, This is the signal after adding the i-th white noise, i.e., the historical power signal of the wind power plant with added components; This is the original signal, i.e., the historical power signal of the target wind power plant; This is the signal-to-noise ratio control coefficient corresponding to the first IMF component of the white noise; For white noise The first IMF component (i.e., intrinsic mode function component) obtained by EMD. Let be the i-th white noise.

[0084] Furthermore, for all signals with different noise added... Calculate their local means; then average all local means to reduce the bias caused by individual noise samples, and obtain the average local mean; subtract this average local mean from the historical power signal of the target wind farm to obtain the first IMF (Intrinsic Mode Function Component) of the historical power signal of the target wind farm. This step stabilizes the white noise IMF through "noise disturbance + averaging". White noise extraction:

[0085] ;

[0086] ;

[0087] in, This is the local mean obtained after the first mean calculation; To add the total number of historical power signals from wind power plants; To calculate the local mean of the signal (usually obtained using EMD). Add component to the historical power signal of the i-th wind power plant The local mean; The first intrinsic mode function component of the power signal of the target wind power plant; The target is the power signal of the wind power plant.

[0088] Furthermore, the average local mean is used as the input for the next algorithm, and the above operation is repeated. That is, the average local mean is used as the new historical power signal of the target wind farm, and new high-frequency modes are extracted for each white noise. (The new intrinsic mode function component, i.e., the second IMF), that is, extracting the (L+1)th intrinsic mode function component to continue approximating the low-frequency structure of the original signal:

[0089] ;

[0090] ;

[0091] in, This is the local mean obtained after the second mean calculation; For white noise The second IMF component obtained by EMD; This is the signal-to-noise ratio control coefficient corresponding to the second IMF component of the white noise; The second intrinsic mode function component of the target wind power plant power signal.

[0092] Furthermore, the above reasoning process is recursively applied. For each round of residuals, white noise representing the k-th IMF is added, and the local mean is calculated and averaged to obtain the IMF. This process is iterated until the obtained average local mean cannot be further decomposed, i.e., the average local mean becomes the trend term.

[0093] ;

[0094] ;

[0095] in, This is the average local mean obtained after calculating the k-th mean. This is the local mean obtained after calculating the mean for the (k-1)th time. This is the signal-to-noise ratio control coefficient corresponding to the k-th IMF component; For white noise The k-th IMF component obtained by EMD; The k-th intrinsic mode function component of the target wind power plant power signal.

[0096] In this embodiment, the Lth intrinsic mode function (IMF) component of each white noise is extracted, and the Lth IMF component of each white noise is added to the historical power signal of the target wind power plant, outputting multiple added component historical power signals of the wind power plant; the local mean of the historical power signal of the wind power plant for each added component is calculated, and the local mean of the historical power signal of the wind power plant for each added component is averaged to output the average local mean; the historical power signal of the target wind power plant is subtracted from the average local mean to output the Lth IMF component of the power signal of the target wind power plant; the average local mean is then determined. The system determines whether the value becomes a trend term. If so, the corresponding local mean becomes the residual term. If not, the (L+1)th intrinsic mode function (IMF) component is extracted from each white noise, and the local mean is used as the new historical power signal of the target wind farm. The (L+1)th IMF component of each white noise is added to the new historical power signal of the target wind farm, outputting multiple new added components of the historical power signal of the wind farm. The process of calculating the local mean and subtracting is repeated, outputting the (L+1)th IMF component of the target wind farm power signal, until the local mean becomes the trend term. To enhance the model's adaptability to sequence complexity, this invention uses the SE index to evaluate the regularity and information content of each IMF component. A lower SE value indicates a stable subsequence structure and strong predictability, and such components are preferentially retained for reconstruction. SE calculation includes multiple processes such as constructing embedding vectors, calculating the maximum difference distance, counting the number of similar vector pairs, and taking the logarithmic ratio. Finally, representative IMFs and residual terms are combined to restore the structurally optimized input data sequence.

[0097] Step 103: Reconstruct the power prediction sequence based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant.

[0098] Preset embedding dimension refers to the preset dimension parameter in phase space reconstruction, which is used to convert a one-dimensional time series into a vector sequence in a high-dimensional space to better capture the dynamic evolution characteristics of the sequence.

[0099] The residual term is the component remaining after mode decomposition, which reflects the overall trend of the historical power signal of the target wind power plant (such as long-term stability or slow rise and fall), and cannot be further decomposed into intrinsic mode functions.

[0100] Power prediction sequences are high-dimensional time series sequences that contain multi-scale features and trend information of the original power signals after phase space reconstruction and integration. They serve as an intermediate carrier connecting mode decomposition and model prediction, providing comprehensive input features for the prediction model.

[0101] Specifically, step 103 may include the following sub-steps:

[0102] S31. Based on the preset embedding dimension, construct multiple embedding vectors for each intrinsic mode function component corresponding to the historical power signal of the target wind power plant.

[0103] S32. Calculate the distance similarity between each embedding vector of each intrinsic mode function component;

[0104] S33. Count the number of embedding vectors corresponding to distance similarity less than the preset similarity threshold, and obtain the local similarity count corresponding to each embedding vector of each intrinsic mode function component.

[0105] S34. Calculate the local count of each intrinsic mode function component based on the local similarity count corresponding to each embedding vector of each intrinsic mode function component.

[0106] S35. Update the preset embedding dimension and determine the new preset embedding dimension;

[0107] S36. Based on the new preset embedding dimension, determine the new local count of each intrinsic mode function component;

[0108] S37. Calculate the sample entropy of each intrinsic mode function component based on the new local counts and local counts of each intrinsic mode function component.

[0109] S38. Compare the sample entropy of each intrinsic mode function component with the preset sample entropy threshold respectively;

[0110] S39. Take the intrinsic mode function component corresponding to any sample entropy that is less than or equal to the preset sample entropy threshold as the target intrinsic mode function component.

[0111] S40. Reconstruct the power prediction sequence based on the target intrinsic mode function components and residual terms.

[0112] Sample entropy (SE) is a widely used method for evaluating the complexity and regularity of time series data, improving upon the approximate entropy (ApEn) method. Compared to ApEn, sample entropy performs better when handling short time series or noisy data. The core idea of ​​sample entropy is to assess the similarity between subsequences in a time series: a lower SE value indicates more similar patterns and a more regular sequence; conversely, a higher SE value indicates a more complex and less regular sequence. The calculation of sample entropy depends on three key parameters: the length of the time series, the similarity tolerance threshold, and the embedding dimension.

[0113] It should be noted that the original time series is constructed into an embedding vector with a predefined embedding dimension of m, resulting in N-m+1 vectors. That is, based on the preset embedding dimension, multiple embedding vectors are constructed for each intrinsic mode function component corresponding to the historical power signal of the target wind farm. The purpose of this step is to transform the time series into adjacent state space vectors, thereby comparing the similarity of wind farm power in multiple dimensions, which is beneficial for comparison. The embedding vector can be represented as:

[0114] ;

[0115] in, Let i be the i-th embedding vector; For the first embedding vector Each element.

[0116] Furthermore, the similarity of the historical power time series between the embedding vectors is calculated:

[0117] ;

[0118] in, To determine the embedding vector when the preset embedding dimension is m. and Distance similarity is used to measure the similarity between two embedding vectors in historical power time series; the larger the distance, the lower the similarity. Embedded vector The Middle Individual elements (power values); Embedded vector The Middle Individual elements (power values); This is the offset index of the element within the embedded vector.

[0119] Furthermore, the distance between two vectors is calculated based on the maximum absolute difference distance. This distance captures the dimension of the largest deviation between the two vectors. Choosing the maximum difference as the distance helps to quickly identify mismatches in high-dimensional space. The number of vector pairs that satisfy the similarity threshold (the ratio of all embedded vectors whose distance is less than the threshold r) is counted and normalized, i.e., the number of embedded vectors corresponding to distance similarity less than the preset similarity threshold is counted. This yields the local similarity count for each embedded vector of each intrinsic mode function component. Based on the local similarity count for each embedded vector of each intrinsic mode function component, the local count for each intrinsic mode function component is calculated.

[0120] ;

[0121] in, This is the local count of similar patterns between embedded vectors in the historical power time series, i.e., the local count of intrinsic mode function components, when the embedding dimension m and the similarity tolerance threshold r are preset. It reflects the overall frequency of occurrence of "similar patterns" under this dimension and threshold. The total length of the historical power time series (total number of data points) is the basic sample size for calculating the statistics; is the local similarity count of the i-th embedding vector in embedding dimension m, where the distance between it and other embedding vectors is less than the similarity threshold r. It is the statistical value of the similarity pattern of a single embedding vector; r is the similarity tolerance threshold, which reflects the overall similarity of the power subsequence in the preset embedding dimension m.

[0122] Furthermore, the embedding dimension is increased to m+1, that is, the preset embedding dimension is updated to determine a new preset embedding dimension m+1. The above steps are repeated to obtain a new local count of the intrinsic mode function components. , representing the local count of similar patterns between embedded vectors in the historical power time series when the new preset embedding dimension m+1 and the similarity tolerance threshold r are used, and finally taking the logarithmic ratio, yields the sample entropy, which represents the decrease in the probability of maintaining similarity after increasing the dimension by one position. The calculation process of sample entropy can be expressed as:

[0123] ;

[0124] in, The sample entropy is used to measure the complexity and regularity of historical power time series under a preset embedding dimension m and similarity tolerance threshold r. The smaller the value, the stronger the regularity of the sequence. For the intrinsic mode function components in the preset embedding dimension, The local count at that time is the new local count of the intrinsic mode function components; For the intrinsic mode function components in the preset embedding dimension, Local counts at time, i.e., local counts of eigenmode function components; Preset embedding dimension; This refers to the updated preset embedding dimension, i.e., the new preset embedding dimension. The length of the time series; This is the similarity tolerance threshold.

[0125] Furthermore, the sample entropy of each intrinsic mode function (IMF) component is compared with a preset sample entropy threshold; the IMF component corresponding to any sample entropy less than or equal to the preset sample entropy threshold is taken as the target IMF component; based on the target IMF component and the residual term, the power prediction sequence is reconstructed. That is, after the above analysis, a set of target IMFs and a residual term are finally obtained as the long-term trend part of the signal. The original signal can be reconstructed by the following formula:

[0126] ;

[0127] in, For power prediction sequences; The total number of the target intrinsic mode function components; For the k-th target intrinsic mode function component; This is the residual term.

[0128] In this embodiment, based on a preset embedding dimension, the one-dimensional time-series data of each intrinsic mode function component corresponding to the historical power signal of the target wind power plant is converted into multiple high-dimensional embedding vectors through a sliding window (each vector contains consecutive data points of the preset embedding dimension). Then, the distance between any two embedding vectors in each intrinsic mode function component is calculated to measure their similarity. The number of embedding vectors whose distance to other embedding vectors is less than a preset similarity threshold is counted to obtain the local similarity count corresponding to each embedding vector. The average of the local similarity counts of all embedding vectors of each intrinsic mode function component is taken as the local count of that component. Then, the preset embedding dimension is updated to the original dimension plus 1 (i.e., the new preset embedding dimension). Based on the new dimension, the process of constructing embedding vectors, calculating distance similarity, counting local similarities, and averaging is repeated to obtain new local counts for each intrinsic mode function (IMF) component. Then, the logarithm of the ratio of the new local count to the original local count is used to calculate the sample entropy of each IMF component. Each sample entropy is compared with a preset sample entropy threshold, and IMF components with sample entropy less than or equal to the threshold are selected as target IMF components (these components have stronger regularity). Finally, all target IMF components and residual terms are aligned and integrated according to time index to form a complete power prediction sequence that preserves the key fluctuation patterns and overall trends of the original power signal, laying the foundation for subsequent input into the target power prediction model.

[0129] Step 104: Based on the input variable characteristics of the historical power signal of the wind power plant, determine the input space sub-region to which the historical power signal of the wind power plant belongs.

[0130] It should be noted that this invention divides sub-regions based on the spatial characteristics of input variables; specifically, it determines the input spatial sub-region to which the historical power signals of wind power plants belong based on the input variable characteristics. For each sub-region, the optimal prediction model is selected by evaluating the training error. For new samples, the system classifies them into the corresponding sub-region based on their characteristics and calls the best model to output the predicted value. This strategy effectively improves the overall prediction accuracy through a model-space matching mechanism.

[0131] Specifically, when determining the input space sub-region to which a wind power plant belongs based on the input variable characteristics of its historical power signals, the input variable features reflecting its key characteristics (such as the mean, standard deviation, peak value, periodicity, and associated meteorological features of the power sequence) are first extracted from the historical power signals. These features are then used as high-dimensional coordinates to construct the input space. Subsequently, the input space is divided into several sub-regions by a preset partitioning rule (such as clustering algorithms based on feature similarity or feature threshold interval partitioning). Each sub-region corresponds to a set of historical power signals with similar feature patterns. Finally, the currently extracted input variable features are matched with the feature range of each sub-region to determine the specific input space sub-region to which the historical power signal belongs. Among them, "input variable features" refers to quantitative indicators (such as statistical features, trend features, fluctuation features, etc.) extracted from the historical power signals of wind power plants that can reflect the data patterns. These are the core parameters describing the characteristics of the power signals. "Input space" is a high-dimensional space composed of all possible combinations of input variable features, with each historical power signal sample corresponding to a point in the space. "Input space sub-regions" are local regions in the input space divided according to feature similarity. Historical power signals within the same sub-region have similar variation patterns, providing a basis for subsequent matching and adaptation of prediction models.

[0132] Step 105: Select the target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the wind power plant power prediction value.

[0133] It should be noted that this invention employs four different recurrent neural network models to learn the reconstructed power sequence:

[0134] Model 1: LSTM-RNN (Long Short-Term Memory Recurrent Neural Network), which solves the gradient vanishing problem through a gating mechanism and is suitable for modeling long-term dependency information; LSTM-RNN is a special type of recurrent neural network (RNN) designed to learn long-term dependency information. It solves the common gradient vanishing and exploding problems in RNNs by introducing forget gates, input gates, and output gates to control the flow of historical information. The LSTM-RNN model can be represented as:

[0135] ;

[0136] in, These represent the forget gate, input gate, and output gate, respectively. This represents the sigmoid activation function; These represent the weight matrices between the forget gate, input gate, generator, and output gate, respectively. This indicates the hidden state of the previous time step; The input value represents the current time step; These represent the bias terms for the forget gate, input gate, generator, and output gate, respectively. Let be the candidate cell state at time step t; Let be the cell state at time step t; Let be the hidden state at time step t.

[0137] Model 2: BiLSTM-RNN (Bidirectional Long Short-Term Memory Recurrent Neural Network), which adds bidirectional networks before and after LSTM to enhance the learning of global sequence context. BiLSTM-RNN is an improved version of LSTM-RNN, capturing the bidirectional contextual dependencies of the sequence by simultaneously using forward and backward LSTM layers. This structure allows BiLSTM-RNN to better understand the overall structure of the sequence, thereby improving prediction accuracy. The advantage of BiLSTM-RNN lies in its ability to utilize both past and future information, resulting in a more comprehensive understanding of the sequence. The BiLSTM-RNN model can be represented as:

[0138] ;

[0139] in, These represent the hidden states of the forward and backward LSTM layers at time step t, respectively. These represent the weight matrices for the forward and backward LSTM layers, respectively. These represent the bias terms for the forward and backward LSTM layers, respectively.

[0140] Model 3: GRU-RNN (Gated Recurrent Unit Recurrent Neural Network), with a simplified structure and higher computational efficiency, is suitable for fast prediction scenarios. GRU-RNN is a variant of RNN that simplifies the structure of LSTM-RNN and improves efficiency by introducing update and reset gates to control the information flow. The GRU-RNN model can be represented as:

[0141] ;

[0142] in, These represent updating the door and resetting the door, respectively. This represents element-wise multiplication.

[0143] Model 4: BiGRU-RNN (Bidirectional Gated Recurrent Unit Recurrent Neural Network), which combines GRU and bidirectional structure, balancing efficiency and context modeling capabilities. BiGRU-RNN is an improved version of GRU-RNN, capturing bidirectional contextual dependencies in sequences by simultaneously using forward and backward GRU layers. The advantage of BiGRU-RNN is its ability to utilize both past and future information, achieving similar results to BiLSTM-RNN but with higher computational efficiency. The BiGRU-RNN model can be represented as:

[0144] ;

[0145] in, These represent the hidden states of the forward and backward GRU layers at time step t, respectively.

[0146] Furthermore, within the input variable space of wind power plant power data, different prediction models exhibit varying predictive performance across different regions. To fully leverage this variation, this strategy divides the input space into several sub-regions and selects the model with the smallest prediction error within each region as the optimal predictor for that region. During the training phase, a comprehensive analysis of the training set determines the optimal model for each sub-region, thereby achieving optimal matching between the region and the most accurate predictor, thus improving overall prediction accuracy.

[0147] For a given time series S, its optimal prediction model (target prediction model) Determined by the following formula:

[0148] ;

[0149] in, Represents the i-th model pair sequence The predicted value, Z represents the actual observed values; Z is a sub-region of the input space, representing a set of power signals from wind power plants with similar characteristics; N is the total number of models participating in model selection.

[0150] Once spatial partitioning is complete, new time-series samples can be categorized into a specific region based on their characteristics, and the optimal model corresponding to that region can be used for prediction. Prediction results (predicted power output from wind power plants). It is expressed as follows:

[0151] ;

[0152] in, The optimal model index corresponding to the input space region Z The optimal model is determined and outputs a function to predict samples within the input space region Z. The advantage of this strategy lies in its ability to dynamically select the most advantageous model based on data characteristics, thereby improving overall prediction accuracy. This method mainly consists of two stages: the training stage completes the partitioning of the input space and the selection of the optimal model; the testing stage selects the appropriate model for prediction output based on the input samples.

[0153] Specifically, when selecting the target power prediction model in this input space sub-region, since the prediction errors of multiple candidate prediction models (such as LSTM, BiGRU, etc.) have been evaluated in advance using training data for each sub-region, the model with the smallest prediction error determined through evaluation in that sub-region is directly called as the target power prediction model. Finally, the power prediction sequence obtained in the previous processing, which contains key information about the target intrinsic mode function components and residual terms, is input into the target power prediction model. The model uses its adaptability to the signal patterns of this sub-region to perform calculations, and finally outputs a wind power prediction value that highly matches the actual power change trend.

[0154] In this invention, the optimal prediction model is dynamically selected based on a classification strategy:

[0155] 1) Divide the input space into sub-regions and calculate the prediction error of each model in each sub-region of the training set; 2) Assign the model with the smallest error to each sub-region as the optimal predictor; 3) For a new sample, call the corresponding optimal model to output the predicted value based on the sub-region to which its features belong. .

[0156] As a comparison of technical effectiveness, existing technologies can be used as a reference. Current prediction technologies have limited capabilities in quantitatively predicting abnormal power fluctuations (such as holiday effects and the impact of extreme weather), often only able to make qualitative judgments and unable to provide precise information on the magnitude of anomalies. This limits their in-depth application in power system planning and management. Although some studies have attempted to use deep learning methods to handle the complexity of power data, these models either fail to fully explore the multi-scale features in the data or lack refined consideration of model applicability. As a result, there is still room for improvement in capturing long-term dependencies and responding to data mutations, especially in areas where different prediction scenarios need to be distinguished to improve overall accuracy.

[0157] To address the aforementioned problems, this invention proposes a wind power prediction method to solve the technical problem of low prediction accuracy caused by insufficient processing of non-stationary signals in existing wind power prediction methods. For example... Figure 2As shown, the specific steps include: acquiring wind power plant power data, preprocessing it, and performing Z-SCORE normalization; using an improved adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) technique to decompose the power data into multiple intrinsic mode function (IMF) components and residual terms; calculating the sample entropy (SE) value of each IMF component and selecting key IMF components based on the sample entropy; constructing four types of prediction models: LSTM-RNN, BiLSTM-RNN, GRU-RNN, and BiGRU-RNN; implementing a classification-based prediction strategy: dividing the input space into regions and assigning the optimal model with the smallest prediction error to each region; training the optimal model for each region using the selected IMF components to reconstruct the power prediction results; and outputting the final wind power plant power prediction value through a dynamic model selection mechanism.

[0158] In summary, this invention fully integrates noise robustness enhancement processing, complexity adaptive modeling, and dynamic model selection mechanisms, enabling high-precision prediction even under challenging conditions such as multi-source noise and irregular missing power data, making it suitable for demanding scenarios such as smart grids.

[0159] In this embodiment of the invention, a method for predicting the power of a wind power plant is provided. The method acquires historical power signals from the wind power plant, preprocesses these signals, and outputs the historical power signal of the target wind power plant. An improved adaptive noise-complete set empirical mode decomposition method is used to generate residual terms and multiple intrinsic mode function (IMF) components corresponding to the historical power signal of the target wind power plant based on the historical power signal and multiple white noises. A power prediction sequence is reconstructed based on a preset embedding dimension and the residual terms and IMF components corresponding to the historical power signal of the target wind power plant. Finally, based on the input variable characteristics of the historical power signal of the wind power plant, the method determines the historical power information of the wind power plant. The input space sub-region to which the signal belongs; within the input space sub-region, a target power prediction model is selected, and the power prediction sequence is input into the target power prediction model for prediction, outputting the wind power plant power prediction value; based on the above scheme, this invention optimizes the noise injection strategy and mode screening mechanism through an improved decomposition method, which can effectively suppress the mode aliasing phenomenon commonly found in traditional mode decomposition, while enhancing the thoroughness of decomposition of weak non-stationary components, so that the reconstructed power prediction sequence completely retains the dynamic law and detailed features of the original signal; combined with the input space sub-region division and target model selection based on historical power features, the prediction model can be specifically adapted to power sequences with different laws, thereby improving the power prediction accuracy.

[0160] Please see Figure 3 , Figure 3 This is a structural block diagram of a wind power plant power prediction system provided in Embodiment 2 of the present invention.

[0161] This invention provides a power prediction system for wind power plants, comprising:

[0162] The acquisition module 301 is used to acquire the historical power signal of the wind power plant, preprocess the historical power signal of the wind power plant, and output the historical power signal of the target wind power plant.

[0163] The generation module 302 is used to generate the residual term and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises using an improved adaptive noise complete set empirical mode decomposition method.

[0164] Reconstruction module 303 is used to reconstruct the power prediction sequence based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant.

[0165] The determination module 304 is used to determine the input spatial sub-region to which the historical power signal of the wind power plant belongs based on the input variable characteristics of the historical power signal of the wind power plant.

[0166] The prediction module 305 is used to select a target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the power prediction value of the wind power plant.

[0167] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0168] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the wind power plant power prediction method as described in any of the above embodiments.

[0169] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implement the steps of the wind power plant power prediction method as described in any of the above embodiments.

[0170] This invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the wind power plant power prediction method as described in any of the above embodiments.

[0171] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0172] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0173] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the power output of a wind power plant, characterized in that, include: The historical power signal of the wind power plant is acquired, and the historical power signal of the wind power plant is preprocessed to output the historical power signal of the target wind power plant. An improved adaptive noise complete set empirical mode decomposition method is adopted to generate residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises. Based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant, the power prediction sequence is reconstructed. Based on the input variable characteristics of the historical power signal of the wind power plant, the input space sub-region to which the historical power signal of the wind power plant belongs is determined; Select a target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the wind power plant power prediction value.

2. The wind power plant power prediction method according to claim 1, characterized in that, The step of preprocessing the historical power signal of the wind power plant and outputting the historical power signal of the target wind power plant includes: The historical power signals of the wind power plants are processed to remove outliers and complete the data, and the intermediate historical power signals of the wind power plants are output. The historical power signals of the intermediate wind power plants are normalized, and the power signals of the target wind power plants are output.

3. The wind power plant power prediction method according to claim 1, characterized in that, The improved adaptive noise-complete set empirical mode decomposition method generates residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal and multiple white noises, including: Extract the intrinsic mode function component of each of the white noises, and add the intrinsic mode function components of each of the white noises to the historical power signal of the target wind power plant, and output multiple historical power signals of the wind power plant with added components; Calculate the local mean of the historical power signal of each of the added component wind power plants, and perform mean operation on the local mean of the historical power signal of each of the added component wind power plants to output the average local mean. The intrinsic mode function components of the target wind power plant's power signal are output by subtracting the historical power signal of the target wind power plant from the average local mean. Determine whether the local mean becomes a trend term; If so, the average local mean corresponding to the trend term will be used as the residual term; If not, extract new intrinsic mode function components from each of the white noises and use the average local mean as the new historical power signal of the target wind farm. The process jumps to the step of adding the intrinsic mode function components of each of the white noises to the historical power signal of the target wind power plant, and outputting multiple historical power signals of the wind power plant with added components, until the average local mean becomes the trend term.

4. The wind power plant power prediction method according to claim 1, characterized in that, The process of reconstructing the power prediction sequence based on a preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant includes: Based on the preset embedding dimension, multiple embedding vectors are constructed for each intrinsic mode function component corresponding to the historical power signal of the target wind power plant. Calculate the distance similarity between each embedding vector of each intrinsic mode function component; The number of embedding vectors corresponding to distance similarities less than a preset similarity threshold is counted to obtain the local similarity counts corresponding to each embedding vector of each intrinsic mode function component. The local count of each intrinsic mode function component is calculated based on the local similarity count corresponding to each embedding vector of each intrinsic mode function component. The preset embedding dimension is updated to determine a new preset embedding dimension; Based on the new preset embedding dimension, a new local count is determined for each of the intrinsic mode function components; Calculate the sample entropy of each intrinsic mode function component based on the new local count and local count of each intrinsic mode function component; The sample entropy of each intrinsic mode function component is compared with a preset sample entropy threshold. The intrinsic mode function component corresponding to any sample entropy less than or equal to the preset sample entropy threshold is taken as the target intrinsic mode function component. The power prediction sequence is reconstructed based on the target intrinsic mode function components and the residual terms.

5. The wind power plant power prediction method according to claim 4, characterized in that, The formula for calculating the sample entropy is as follows: ； in, The sample entropy; For the intrinsic mode function components in the preset embedding dimension, The local count at that time is the new local count of the intrinsic mode function components; For the intrinsic mode function components in the preset embedding dimension, Local counts at time, i.e., local counts of eigenmode function components; Preset embedding dimension; This refers to the updated preset embedding dimension, i.e., the new preset embedding dimension. The length of the time series; This is the similarity tolerance threshold.

6. The wind power plant power prediction method according to claim 4, characterized in that, The power prediction sequence is specifically as follows: ； in, For power prediction sequences; The total number of the target intrinsic mode function components; For the k-th target intrinsic mode function component; This is the residual term.

7. A power prediction system for wind power plants, characterized in that, include: The acquisition module is used to acquire historical power signals of wind power plants, preprocess the historical power signals of wind power plants, and output the historical power signals of target wind power plants. The generation module is used to generate residual terms and multiple intrinsic mode function components corresponding to the historical power signal of the target wind power plant based on the historical power signal of the target wind power plant and multiple white noises using an improved adaptive noise complete set empirical mode decomposition method. The reconstruction module is used to reconstruct the power prediction sequence based on the preset embedding dimension and the residual terms and multiple intrinsic mode function components corresponding to the historical power signals of the target wind power plant. The determination module is used to determine the input spatial sub-region to which the historical power signal of the wind power plant belongs based on the input variable characteristics of the historical power signal of the wind power plant. The prediction module is used to select a target power prediction model in the input space sub-region, input the power prediction sequence into the target power prediction model for prediction, and output the power prediction value of the wind power plant.

8. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the wind power plant power prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the wind power plant power prediction method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the wind power plant power prediction method as described in any one of claims 1-6.

Images