Feature construction method for online wind power prediction

Abstract

This application provides a feature construction method for online wind power prediction. By constructing a sliding historical window constrained by the prediction time, the feature construction process relies solely on historical data, meeting the temporal requirements of online prediction scenarios. Furthermore, by establishing initial parameter states and inheriting parameter states between windows, the variational mode decomposition results of adjacent windows remain continuous, reducing jumps in the number and frequency distribution of modal components. Through multi-dimensional initial decomposition quality assessment and a triggered local parameter update mechanism, parameters are only locally adjusted when decomposition quality deteriorates or anomalies occur in a certain dimension, balancing decomposition adaptability with online computational complexity. By performing modal channel alignment, channel merging, and zero-padding operations based on the center frequency, a fixed-dimensional multi-scale feature matrix is ​​generated. Validity markers and abrupt window markers are introduced as auxiliary features, ensuring the output features remain stable in terms of channel meaning, dimension, and temporal structure.

Description

Technical Field

[0001] This application relates to the field of signal processing, and in particular to a feature construction method for online wind power prediction. Background Technology

[0002] Wind power exhibits strong randomness, volatility, and non-stationarity, with its variations influenced by factors such as wind speed, wind direction, meteorological conditions, and the operating status of wind turbines. In wind power forecasting, directly inputting the raw power sequence into the prediction model often fails to effectively extract trend changes, fluctuations, and high-frequency disturbances across different time scales. Therefore, engineering applications and research typically incorporate a feature construction stage at the front end of the prediction model to improve the expressiveness of the input data.

[0003] In some feasible implementations, a common approach is to use Variational Mode Decomposition (VMD) to decompose wind power sequences or sliding window samples, breaking down the original power sequence into several Intrinsic Mode Functions (IMFs). These IMFs are then used as multi-scale features input into subsequent prediction models. This method can reduce the non-stationarity of the original wind power sequence to some extent, enabling the prediction model to obtain power variation information at different frequency scales. In practical applications, existing methods typically determine the number of modes and penalty factor required for VMD decomposition before decomposing the training samples or prediction windows. Some methods use fixed VMD parameters throughout the prediction process to reduce computational load; others adjust parameters for different sliding windows to improve adaptability; still others directly input the multiple IMFs obtained from different window decompositions into the prediction model according to their original indices.

[0004] However, existing feature construction methods based on variational mode decomposition lack clear definition of the time boundary between historical input windows and future prediction targets. Some construction processes may include data after the prediction target time in the decomposition, making it difficult to meet online prediction requirements. Fixed-parameter methods struggle to adapt to changes in wind power operating conditions; window-by-window parameter optimization methods involve significant online computation. Insufficient consideration of parameter continuity between adjacent windows can easily lead to jumps in the number of modal components, center frequencies, and energy distribution. Furthermore, directly organizing features according to decomposition indices may result in different frequency scales for components with the same indices in different windows, leading to inconsistent meanings of feature channels.

[0005] In summary, current online wind power prediction methods based on variational mode decomposition have limitations in terms of data validity and computational complexity. A feature construction method for online wind power prediction is needed to achieve continuous, stable, and fixed-dimensional VMD feature construction, thereby improving data validity and reducing computational complexity. Summary of the Invention

[0006] The purpose of this application is to at least address one of the aforementioned technical deficiencies, particularly the technical limitations of existing online wind power prediction methods based on variational mode decomposition, which suffer from insufficient feature construction effectiveness, resulting in limitations in data validity and computational complexity.

[0007] Firstly, this application provides a feature construction method for online wind power prediction, the method comprising: Obtain the historical wind power sequence of the target wind farm; Based on the window parameters, a sliding historical window corresponding to the prediction time is constructed based on the historical wind power sequence. The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. For the sliding history window at the current moment, based on the final parameter state of the sliding history window at the previous moment, variational mode decomposition corresponding to the prediction moment is performed to obtain multiple target mode components; Each target modal component is mapped to a fixed output channel according to its corresponding center frequency to generate a fixed-dimensional multi-scale feature.

[0008] As an optional implementation, the window parameters include historical window length, prediction step size, and sliding step size; The acquisition of the historical wind power sequence of the target wind farm includes: Obtain the raw wind power data of the target wind farm; The original wind power data is time-aligned and resampled according to a preset sampling interval, and the sampling time corresponding to each sampling point is adjusted to obtain resampled wind power data. The resampled wind power data is validated, abnormal data is removed, and missing data is filled in by interpolation or previous values ​​to form a historical sequence of wind power. And, determine the window parameters.

[0009] As an optional implementation, constructing a sliding historical window corresponding to the prediction time based on the historical wind power sequence according to window parameters includes: Based on the sliding step size, determine the prediction anchor points corresponding to multiple prediction times; Based on the historical window length, the sliding historical window is formed by taking the prediction anchor point and a number of consecutive historical sampling points before it. Wherein, the prediction target corresponding to the sliding history window is the power value of the target sampling point corresponding to the prediction anchor point; The prediction anchor point is moved along the time axis according to the sliding step size to obtain the next sliding history window; Adjacent sliding history windows share multiple sampling points based on the history window length and the sliding step size; Furthermore, the method further includes: Output the window index information corresponding to each of the aforementioned sliding history windows; The window index information is used to indicate the attribute information and order relationship of each sliding history window, including the current window number, window start time, window end time, prediction target time, prediction step size, sliding step size, window quality mark and the previous sliding window number. The window index information is associated with and stored in relation to the historical power vector corresponding to the sliding history window.

[0010] As an optional implementation, the method further includes: The historical power vectors corresponding to each of the sliding historical windows are arranged according to the order of each of the sliding historical windows to form a two-dimensional window matrix; Based on the two-dimensional window matrix, select the first preset number of sliding history windows as the initial parameter calibration sample group; Construct a candidate set of modality numbers and a candidate set of penalty factors to obtain multiple candidate combinations of window parameters; Wherein, the candidate set of modality numbers is a preset set of discrete integers, and the candidate set of penalty factors is set according to linear intervals or logarithmic intervals; For each sliding history window in the initial parameter calibration sample group, variational mode decomposition is performed under each candidate combination of window parameters to obtain candidate mode components and corresponding center frequencies; A comprehensive score corresponding to the candidate combination of the window parameters is calculated based on at least two of the following indicators: reconstruction accuracy, modal correlation, frequency separation, and modal redundancy. When a single sliding history window is used, the corresponding comprehensive score is directly used as the final score. When multiple sliding history windows are used, the average or median of the comprehensive scores of each sliding history window is used as the final score. The candidate combination of window parameters with the best final score is selected as the initial window parameters. Based on the initial window parameters, the initial variational mode decomposition is performed on the target sliding window to obtain multiple initial component channels. The channel order, center frequency and energy ratio of each initial component channel are determined. The target sliding window includes the first sliding history window in the initial parameter calibration sample group in chronological order; The initial window parameters, the final score, the window number corresponding to the target sliding window, the channel order, the center frequency, and the energy percentage are used as the initial parameter states corresponding to the prediction time.

[0011] As an optional implementation, after performing variational mode decomposition corresponding to the prediction time on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, to obtain multiple target mode components, the method further includes: Calculate the initial decomposition quality assessment value corresponding to the current moment; The initial decomposition quality assessment value is obtained by weighted calculation based on multiple preset dimensions. The preset dimensions include the reconstruction accuracy, modal correlation, frequency separation, modal redundancy, center frequency drift and energy ratio drift between the sliding history window at the current time and the sliding history window at the previous time for each target modal component. When the initial decomposition quality assessment value does not meet the preset quality threshold, or when the value of any preset dimension exceeds the preset dimension abnormality threshold, the local parameters of the current sliding history window are updated. When the initial decomposition quality assessment value meets the preset quality threshold, and the values ​​of all preset dimensions do not exceed the preset dimension abnormality threshold, the final parameter state inherited from the previous sliding history window is retained.

[0012] As an optional implementation, updating the local parameters of the current sliding history window includes: Based on the neighborhood of the final parameter state of the sliding history window at the previous moment, a local parameter candidate set is constructed. The local parameter combination that optimizes the decomposition quality of the current sliding history window is selected from the local parameter candidate set as the updated parameter state of the current window. Based on the updated parameter state, the initial decomposition parameters inherited by the current sliding history window are locally updated to obtain the updated parameter state. The current sliding history window is then re-executed with variational mode decomposition based on the updated parameter state to obtain the final mode components after the re-execution of variational mode decomposition. The channel order, the center frequency, and the energy percentage are then determined. The updated parameter state, the final score corresponding to the updated parameter state, the window number corresponding to the current sliding history window, the channel order, the center frequency, and the energy percentage are used as the final parameter state corresponding to the current sliding history window.

[0013] As an optional implementation, the step of mapping each of the target modal components to a fixed output channel according to its corresponding center frequency to generate fixed-dimensional multi-scale features includes: Determine the fixed output channel and its corresponding frequency anchor point; For each final modal component obtained by decomposing the current sliding history window, it is mapped to each fixed output channel according to the nearest neighbor relationship or frequency interval affiliation relationship between its center frequency and the frequency anchor point corresponding to each fixed output channel. If multiple final modal components are mapped to the same fixed output channel, then the final modal components are weighted and merged according to energy weights, and the validity flag of the corresponding fixed output channel is set to the first value; If any of the fixed output channels does not have a corresponding final modal component mapping, then the corresponding fixed output channel is padded with zeros, and the validity flag of the corresponding fixed output channel is set to the second value; Based on each of the fixed output channels, determine the fixed-dimensional multi-scale features corresponding to the current sliding history window, and combine them to form a feature matrix corresponding to the current sliding history window; Furthermore, for each of the sliding history windows, based on the center frequency drift, when the difference between the center frequency of the current sliding history window and the center frequency of the corresponding final modal component of the previous sliding history window exceeds a preset drift threshold, and / or when the modal change of the current sliding history window and the corresponding final modal component of the previous sliding history window exceeds a preset modal change threshold, the current sliding history window is determined to be a mutation window, and the validity marker and mutation window marker of each of the fixed output channels in each of the sliding history windows are combined to form an auxiliary feature; Furthermore, the feature matrices corresponding to each of the sliding history windows are combined in chronological order to generate the final output matrix.

[0014] Secondly, this application provides a feature construction device for online wind power prediction, the device comprising: The acquisition module is used to acquire the historical sequence of wind power of the target wind farm; The processing module is used to construct a sliding historical window corresponding to the prediction time based on the wind power historical sequence according to the window parameters; The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. The processing module is further configured to perform variational mode decomposition on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, corresponding to the prediction time, to obtain multiple target mode components. The processing module is further configured to map each of the target modal components to a fixed output channel according to its corresponding center frequency, thereby generating a fixed-dimensional multi-scale feature.

[0015] Thirdly, this application provides a computer device including one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method described in the first aspect.

[0016] Fourthly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method described in the first aspect.

[0017] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: Based on any of the above embodiments, the method provided in this application first constructs a sliding historical window constrained by the prediction time, ensuring that the feature construction process relies solely on historical data, thus meeting the temporal requirements of online prediction scenarios. Building upon this, by establishing initial parameter states and inheriting parameter states between windows, the variational mode decomposition results of adjacent windows remain continuous, reducing jumps in the number and frequency distribution of modal components. Simultaneously, through multi-dimensional initial decomposition quality assessment and a triggered local parameter update mechanism, parameters are locally adjusted only when decomposition quality declines or anomalies occur in a certain dimension, balancing decomposition adaptability with online computational complexity. Finally, through modal channel alignment, channel merging, and zero-padding operations based on center frequency, a fixed-dimensional multi-scale feature matrix is ​​generated, and validity markers and abrupt window markers are introduced as auxiliary features, ensuring the output features remain stable in terms of channel meaning, dimension, and temporal structure. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a feature construction method for online wind power prediction provided in one embodiment of this application; Figure 2 A schematic diagram of the overall process framework of the feature construction method for online wind power prediction provided in one embodiment of this application, based on a real-world application scenario. Figure 3 This is a schematic diagram of a sliding window creation process provided in one embodiment of this application; Figure 4 A schematic diagram of the initial VMD parameter optimization and state caching process provided in one embodiment of this application; Figure 5 A schematic diagram of the initial VMD decomposition process of the current window in the inheritance state provided in one embodiment of this application; Figure 6 This is a schematic diagram of the current window initial decomposition evaluation and local update trigger judgment process provided in one embodiment of this application; Figure 7 This is a schematic diagram of a local candidate parameter update and parameter smoothing process provided in one embodiment of this application; Figure 8 This is a schematic diagram of modal channel alignment and stable multi-scale IMF feature output provided in one embodiment of this application; Figure 9 This is an internal structural diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Wind power exhibits strong randomness, volatility, and non-stationarity, with its variations influenced by factors such as wind speed, wind direction, meteorological conditions, and the operating status of wind turbines. In wind power forecasting, directly inputting the raw power sequence into the prediction model often fails to effectively extract trend changes, fluctuations, and high-frequency disturbances across different time scales. Therefore, engineering applications and research typically incorporate a feature construction stage at the front end of the prediction model to improve the expressiveness of the input data.

[0022] In existing technologies, a common approach is to utilize... The wind power sequence or sliding window sample is decomposed into several parts. Modal components, and then these The components are used as multi-scale features input to the subsequent prediction model. This approach can reduce the non-stationarity of the original wind power sequence to some extent, enabling the prediction model to obtain power variation information at different frequency scales.

[0023] In practical applications, existing methods typically first determine The required number of modalities and penalty factor are decomposed, and then the training samples or prediction window are decomposed. Some methods, to reduce computational cost, use a fixed number of modalities throughout the prediction process. Parameters; some methods adjust parameters separately for different sliding windows to improve adaptability; other methods directly decompose different windows to obtain the parameters. Input the equal components into the prediction model according to their original serial numbers.

[0024] However, these methods have the following limitations: (1) In online prediction scenarios, the input features of each prediction window should only come from historical data before the prediction time. The existing feature construction process does not clearly define the time boundary between the historical input window and the future prediction target, which makes the feature construction process not fully conform to the real online prediction conditions.

[0025] (2) Existing In feature construction methods, using a fixed number of modes and a fixed penalty factor across all windows makes it difficult to adapt to changes in wind power operating conditions. When the frequency structure changes across different windows, fixed parameters may lead to a decrease in decomposition performance.

[0026] (3) If the process is repeated for each sliding window While parameter optimization can improve the adaptability of single-window decomposition, it increases the amount of online computation, which is not conducive to real-time applications in wind power rolling forecasting scenarios.

[0027] (4) Existing methods focus primarily on the decomposition effect of a single window, neglecting the continuity of parameters and decomposition results between adjacent windows. Since parameters between adjacent windows may change frequently, this can easily lead to… The number of components, center frequency, and energy distribution change abruptly.

[0028] (5) Existing methods usually follow directly The serial number organization features, but the same serial number in different windows Components may not correspond to the same frequency scale, which can easily lead to inconsistent meanings of feature channels and affect the subsequent prediction model's learning of temporal features.

[0029] To address these limitations, this application resolves the relevant technical problems through the following means: (1) A sliding window construction method under prediction time constraints is proposed, which makes each window's... Decomposition and feature construction use only historical wind power data prior to the prediction time, avoiding future information leakage and making the feature construction process consistent with real online prediction scenarios.

[0030] (2) A parameter-state inheritance-based method is proposed. The decomposition mechanism enables the current window to inherit state information such as the number of modes, penalty factor, center frequency, and energy distribution of the previous window, thereby enhancing the continuity between adjacent windows.

[0031] (3) A trigger-based local parameter update mechanism is proposed, which enables the system to update local parameters only when the initial decomposition quality of the current window does not meet the requirements, thereby avoiding the need to search for global parameters again for each window and reducing the amount of online computation.

[0032] (4) A parameter smoothing method is proposed to make... The parameters can be adjusted appropriately according to window changes, while avoiding frequent jumps in the number of modalities and penalty factors between adjacent windows.

[0033] (5) A modal channel alignment method is proposed to align the modal channels of different windows. The components are mapped to fixed output channels according to their center frequencies, and then formed into fixed-dimensional multi-scale structures through channel merging and zero-padding of empty channels. Features to solve different windows The issue of inconsistent meanings of the channels.

[0034] Therefore, this invention can achieve continuous, stable, and fixed-dimensional online wind power forecasting while preventing data leakage caused by data decomposition. Feature construction.

[0035] Therefore, the technical concept of this application lies in the fact that the method provided by this application coordinates four levels: sliding window construction, parameter state management, decomposition quality assessment and updating, and modal channel alignment, to achieve continuous, stable, and fixed-dimensional feature construction for online wind power prediction. At the sliding window construction level, the prediction time constraint ensures that each sliding historical window only contains historical power data before the prediction target time, guaranteeing temporal consistency between feature construction and the online prediction scenario from the data source. At the parameter state management level, the initial parameter state is first determined by optimizing a finite candidate set using a pre-set number of windows. Then, through parameter state inheritance between windows, adjacent windows share key information such as the number of modes, penalty factor, center frequency, and energy distribution, enabling most windows to obtain a reasonable decomposition starting point without re-searching for parameters. At the decomposition quality assessment and update level, a comprehensive evaluation is conducted using multiple dimensions, including reconstruction accuracy, modal correlation, frequency separation, modal redundancy, and center frequency and energy percentage drift between adjacent windows. Local parameter updates are triggered only when the quality assessment value fails to meet a threshold or a certain dimension is abnormal, and the update range is limited to the neighborhood of the parameter state of the previous window. This avoids both the insufficient adaptability of fixed parameters and the excessive online computation caused by window-by-window global optimization. At the modal channel alignment level, by mapping the modal components of each window to preset fixed output channels according to their center frequencies, combined with weighted merging and zero-padding of empty channels, different windows always output feature matrices of the same dimension, solving the problem of inconsistent frequency meanings of modal components with the same index in traditional methods. At the same time, validity markers and abrupt window markers for each channel are introduced as auxiliary features, making the feature matrix contain richer state information, which is easier for subsequent prediction models to identify and utilize. The technical features at these levels work together to enable the overall method to output multi-scale features with continuous parameters, fixed dimensions, and consistent channel meanings while meeting the online prediction time constraints, providing stable and reliable feature inputs for subsequent wind power prediction.

[0036] The methods provided in this application will be described in detail below based on the corresponding implementation methods in some practical application scenarios.

[0037] Figure 1 This is a flowchart illustrating a feature construction method for online wind power prediction provided in one embodiment of this application, as shown below. Figure 1 As shown, this application provides a feature construction method for online wind power prediction, the method comprising: S101. Obtain the historical sequence of wind power output of the target wind farm; S102. Based on the window parameters, construct a sliding historical window corresponding to the prediction time according to the historical sequence of wind power; The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. S103. For the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, perform variational mode decomposition corresponding to the prediction time to obtain multiple target mode components. S104. Map each target modal component to a fixed output channel according to its corresponding center frequency to generate a fixed-dimensional multi-scale feature.

[0038] This implementation acquires historical wind power sequences and constructs sliding historical windows based on window parameters. Each sliding historical window contains only historical power data up to and including the prediction time, thus limiting the time range upon which feature construction depends from the data source. This ensures that subsequent decomposition and feature generation processes remain consistent with the temporal constraints of online prediction scenarios. When performing variational mode decomposition on the current sliding historical window, the final parameter state of the previous sliding historical window is directly used as the decomposition starting point, eliminating the need for a new global parameter search. This maintains continuity in the number of modes, center frequencies, and energy distribution between adjacent windows, avoiding mode component jumps caused by frequent parameter changes. After decomposition, each target mode component is mapped to a fixed output channel according to its center frequency, forming multi-scale features with fixed dimensions. This ensures that the feature channels output by different sliding historical windows maintain consistency in frequency meaning, thereby providing structurally stable, temporally continuous, and dimensionally uniform input features for subsequent prediction models.

[0039] As an optional implementation, the window parameters include historical window length, prediction step size, and sliding step size; The acquisition of the historical wind power sequence of the target wind farm includes: Obtain the raw wind power data of the target wind farm; The original wind power data is time-aligned and resampled according to a preset sampling interval, and the sampling time corresponding to each sampling point is adjusted to obtain resampled wind power data. The resampled wind power data is validated, abnormal data is removed, and missing data is filled in by interpolation or previous values ​​to form a historical sequence of wind power. And, determine the window parameters.

[0040] This implementation defines the window parameters as historical window length, prediction step size, and sliding step size, providing a unified parameter benchmark for the construction of subsequent sliding historical windows. This facilitates adjustments to the window size and prediction lead under different prediction scenarios. During the acquisition of historical wind power sequences, the original wind power data is first time-aligned and resampled at preset sampling intervals to ensure consistent intervals between adjacent sampling times, eliminating data inconsistencies caused by uneven acquisition times. The resampled data undergoes validity checks, and outliers are removed. Missing data is interpolated or filled with previous values, ensuring the resulting historical wind power sequence is continuous in time and numerically reasonable. This provides quality-controlled foundational data for subsequent sliding window construction and variational mode decomposition.

[0041] As an optional implementation, constructing a sliding historical window corresponding to the prediction time based on the historical wind power sequence according to window parameters includes: Based on the sliding step size, determine the prediction anchor points corresponding to multiple prediction times; Based on the historical window length, the sliding historical window is formed by taking the prediction anchor point and a number of consecutive historical sampling points before it. Wherein, the prediction target corresponding to the sliding history window is the power value of the target sampling point corresponding to the prediction anchor point; The prediction anchor point is moved along the time axis according to the sliding step size to obtain the next sliding history window; Adjacent sliding history windows share multiple sampling points based on the history window length and the sliding step size; Furthermore, the method further includes: Output the window index information corresponding to each of the aforementioned sliding history windows; The window index information is used to indicate the attribute information and order relationship of each sliding history window, including the current window number, window start time, window end time, prediction target time, prediction step size, sliding step size, window quality mark and the previous sliding window number. The window index information is associated with and stored in relation to the historical power vector corresponding to the sliding history window.

[0042] This implementation determines multiple prediction anchor points based on the sliding step size, and constructs a sliding historical window by taking the prediction anchor points and the historical sampling points preceding them based on the historical window length. This ensures that the prediction target of each window clearly points to the target sampling point after the prediction anchor point, thus establishing a correspondence between historical data and future prediction targets during the window construction stage. Adjacent sliding historical windows share multiple sampling points according to the sliding step size, maintaining a high degree of overlap between windows in the time domain, providing a data continuity basis for the inheritance and local updates of parameter states in subsequent variational mode decomposition. Simultaneously, each sliding historical window outputs window index information containing window number, start and end times, prediction target time, quality flag, and the previous window number, and stores it in association with the historical power vector. This ensures that the sequential relationship and attribute information between windows are completely recorded, providing a clear data organization basis for subsequent parameter state calls and modal channel alignment.

[0043] As an optional implementation, the method further includes: The historical power vectors corresponding to each of the sliding historical windows are arranged according to the order of each of the sliding historical windows to form a two-dimensional window matrix; Based on the two-dimensional window matrix, select the first preset number of sliding history windows as the initial parameter calibration sample group; Construct a candidate set of modality numbers and a candidate set of penalty factors to obtain multiple candidate combinations of window parameters; Wherein, the candidate set of modality numbers is a preset set of discrete integers, and the candidate set of penalty factors is set according to linear intervals or logarithmic intervals; For each sliding history window in the initial parameter calibration sample group, variational mode decomposition is performed under each candidate combination of window parameters to obtain candidate mode components and corresponding center frequencies; A comprehensive score corresponding to the candidate combination of the window parameters is calculated based on at least two of the following indicators: reconstruction accuracy, modal correlation, frequency separation, and modal redundancy. When a single sliding history window is used, the corresponding comprehensive score is directly used as the final score. When multiple sliding history windows are used, the average or median of the comprehensive scores of each sliding history window is used as the final score. The candidate combination of window parameters with the best final score is selected as the initial window parameters. Based on the initial window parameters, the initial variational mode decomposition is performed on the target sliding window to obtain multiple initial component channels. The channel order, center frequency and energy ratio of each initial component channel are determined. The target sliding window includes the first sliding history window in the initial parameter calibration sample group in chronological order; The initial window parameters, the final score, the window number corresponding to the target sliding window, the channel order, the center frequency, and the energy percentage are used as the initial parameter states corresponding to the prediction time.

[0044] This implementation arranges historical power vectors into a two-dimensional window matrix according to the order of sliding historical windows, enabling batch organization of data from multiple windows and facilitating unified management and indexing. A predetermined number of windows are selected from the two-dimensional window matrix as the initial parameter calibration sample group, ensuring that the initial parameters are determined based on actual data rather than empirical settings, thus improving the rationality of parameter initialization. By constructing a discrete candidate set of modal numbers and a candidate set of penalty factors set at intervals, optimization is performed using a finite combination of candidates, avoiding computational fluctuations and result instability caused by random searches. During the scoring process, a comprehensive score is calculated based on at least two of the following indicators: reconstruction accuracy, modal correlation, frequency separation, and modal redundancy, measuring decomposition quality from multiple perspectives and preventing over- or under-decomposition problems caused by a single indicator. Using the average or median of multiple window scores reduces the impact of abnormal fluctuations in a single window on the selection of initial parameters. Finally, the selected initial window parameters and the decomposition results of the target sliding window are completely saved as the initial parameter state, providing a stable and traceable initial benchmark for parameter inheritance and local updates in subsequent windows.

[0045] As an optional implementation, after performing variational mode decomposition corresponding to the prediction time on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, to obtain multiple target mode components, the method further includes: Calculate the initial decomposition quality assessment value corresponding to the current moment; The initial decomposition quality assessment value is obtained by weighted calculation based on multiple preset dimensions. The preset dimensions include the reconstruction accuracy, modal correlation, frequency separation, modal redundancy, center frequency drift and energy ratio drift between the sliding history window at the current time and the sliding history window at the previous time for each target modal component. When the initial decomposition quality assessment value does not meet the preset quality threshold, or when the value of any preset dimension exceeds the preset dimension abnormality threshold, the local parameters of the current sliding history window are updated. When the initial decomposition quality assessment value meets the preset quality threshold, and the values ​​of all preset dimensions do not exceed the preset dimension abnormality threshold, the final parameter state inherited from the previous sliding history window is retained.

[0046] This implementation calculates initial decomposition quality assessment values ​​across multiple preset dimensions, including reconstruction accuracy, modal correlation, frequency separation, modal redundancy, and center frequency and energy percentage drift between adjacent windows. It comprehensively evaluates the initial decomposition quality of the current window from multiple aspects such as reconstruction capability, modal independence, frequency distribution, and temporal continuity, making decomposition quality assessment less reliant on a single indicator and reducing misjudgments caused by fluctuations in individual indicators. Local parameter updates are triggered when the quality assessment value fails to meet a preset quality threshold or when any dimension exceeds an abnormal threshold. This ensures that parameter adjustments are only performed when decomposition quality significantly decreases or anomalies occur in a particular dimension, maintaining decomposition adaptability while reducing unnecessary parameter re-searches. When quality meets the target, inherited parameter states are retained, allowing the parameter states of most windows to continue, thereby maintaining the continuity of parameter states between adjacent windows and the stability of feature construction overall.

[0047] As an optional implementation, updating the local parameters of the current sliding history window includes: Based on the neighborhood of the final parameter state of the sliding history window at the previous moment, a local parameter candidate set is constructed. The local parameter combination that optimizes the decomposition quality of the current sliding history window is selected from the local parameter candidate set as the updated parameter state of the current window. Based on the updated parameter state, local parameter updates are performed on the initial window parameters to obtain the updated parameter state. Variational mode decomposition is then re-executed on the current sliding history window based on the updated parameter state to obtain the final mode components after re-execution of variational mode decomposition, and the channel order, the center frequency, and the energy ratio are determined. The updated parameter state, the final score corresponding to the updated parameter state, the window number corresponding to the current sliding history window, the channel order, the center frequency, and the energy percentage are used as the final parameter state corresponding to the current sliding history window.

[0048] This implementation constructs a local parameter candidate set within the neighborhood of the final parameter state of the previous sliding historical window, limiting the parameter search range to a small area near the existing parameters. This avoids re-optimization across the entire parameter space, thus controlling the online computational load while ensuring decomposition quality. The combination that optimizes the decomposition quality of the current window is selected from the local parameter candidate set as the updated parameter state. This allows the updated parameters to be appropriately adjusted based on the existing parameters, adapting to changes in wind power operation without disrupting the continuity between adjacent windows due to excessive parameter changes. Variational mode decomposition is re-executed based on the updated parameter state to determine the channel order, center frequency, and energy percentage, ensuring that the updated parameter state fully encompasses the decomposition result information of the current window. The updated parameter state, along with the final score, window number, and decomposition result information, is saved as the final parameter state of the current window, providing a complete chain of state information for parameter inheritance in the next window.

[0049] As an optional implementation, the step of mapping each of the target modal components to a fixed output channel according to its corresponding center frequency to generate fixed-dimensional multi-scale features includes: Determine the fixed output channel and its corresponding frequency anchor point; For each final modal component obtained by decomposing the current sliding history window, it is mapped to each fixed output channel according to the nearest neighbor relationship or frequency interval affiliation relationship between its center frequency and the frequency anchor point corresponding to each fixed output channel. If multiple final modal components are mapped to the same fixed output channel, then the final modal components are weighted and merged according to energy weights, and the validity flag of the corresponding fixed output channel is set to the first value; It should be noted that the weighted merging operation can be an equal weight superposition, or an unequal weight superposition that emphasizes some channel information while appropriately ignoring others. In a broader sense, it can also be a direct addition of the final modal components, which can be understood as a proportional amplification of equal weight superposition, and the amplification ratio can be determined according to the actual needs of data processing.

[0050] If any of the fixed output channels does not have a corresponding final modal component mapping, then the corresponding fixed output channel is padded with zeros, and the validity flag of the corresponding fixed output channel is set to the second value; Based on each of the fixed output channels, determine the fixed-dimensional multi-scale features corresponding to the current sliding history window, and combine them to form a feature matrix corresponding to the current sliding history window; Furthermore, for each of the sliding history windows, based on the center frequency drift, when the difference between the center frequency of the current sliding history window and the center frequency of the corresponding final modal component of the previous sliding history window exceeds a preset drift threshold, and / or when the modal change of the current sliding history window and the corresponding final modal component of the previous sliding history window exceeds a preset modal change threshold, the current sliding history window is determined to be a mutation window, and the validity marker and mutation window marker of each of the fixed output channels in each of the sliding history windows are combined to form an auxiliary feature; It should be noted that during the mutation window determination process, the determination can be made based on the center frequency drift or modal change of the corresponding final modal component between the current sliding history window and the previous sliding history window. For details, please refer to step S74 based on the actual application scenario.

[0051] Furthermore, the feature matrices corresponding to each of the sliding history windows are combined in chronological order to generate the final output matrix.

[0052] This implementation predetermines fixed output channels and corresponding frequency anchors, mapping modal components of different sliding historical windows according to a unified frequency reference system, thus ensuring consistency in frequency meaning across output feature channels of each window. Each final modal component is mapped based on the nearest neighbor relationship or frequency interval affiliation between its center frequency and the frequency anchor, ensuring each modal component is assigned to the fixed channel best matching its frequency characteristics. When multiple modal components fall into the same channel, they are weighted and merged according to energy weights; when a channel has no modal components, zeros are padded, ensuring each window outputs a preset number of fixed-dimensional features, thus resolving the issue of feature dimension variations caused by inconsistent modal numbers across different windows. By setting validity markers for each fixed output channel, the actual state of each channel in the feature matrix can be perceived and utilized by subsequent models. Furthermore, by comparing the center frequency difference between the current window and the previous window's corresponding final modal components, windows with differences exceeding a preset drift threshold are marked as abrupt windows. The validity markers of each channel are combined with the abrupt window markers to form auxiliary features, ensuring the feature matrix contains not only time-frequency information of modal components but also marker information for window and channel states. The feature matrices of each window are combined in chronological order to generate the final output matrix, ensuring that the output features remain ordered in the time dimension, thus providing structurally sound and informationally complete input data for subsequent prediction models.

[0053] Please see Figure 2 , Figure 2 This is a schematic diagram of the overall process framework of the feature construction method for online wind power prediction provided in one embodiment of this application, based on a real-world application scenario. The following describes the specific implementation methods of each embodiment of this application in combination with the application scenario.

[0054] like Figure 2 As shown, based on the various embodiments, this application can be further broken down into 7 related steps.

[0055] Specifically, S1: Acquire and preprocess wind power time-series data: Historical active power data of the target wind farm or wind turbine is obtained and processed according to a uniform time interval to form a wind power time series for subsequent sliding window construction. The original data can be represented as:

[0056] In the formula, Indicates the first Each sampling time, This represents the raw wind power value collected at that moment, and the number of raw data points.

[0057] The raw data is sampled at a preset interval. Perform time alignment and resampling to ensure that adjacent sampling times satisfy:

[0058] The resampled data underwent a basic validity check to remove obvious outliers, and a small number of missing data were imputed to obtain a continuous wind power sequence.

[0059] in, Indicates the first The wind power value corresponding to each valid sampling time. This indicates the length of the valid data obtained after preprocessing.

[0060] To avoid abnormal power values ​​affecting subsequent VMD decomposition and feature construction, reasonable value range constraints are set for the power data:

[0061] In the formula, This indicates the rated power of a wind farm or wind turbine.

[0062] After completing the data preprocessing, determine the basic parameters required for constructing the subsequent sliding window, including the history window length. Predicting step size Sliding step size and valid data length The parameters and power sequence mentioned above are used as inputs to step S2 to construct a sliding history window with prediction time constraints.

[0063] Therefore, the output of step S1 is:

[0064] This step is primarily used to ensure the consistency of the input data in terms of time scale and numerical range, laying the foundation for subsequent sliding window construction. Parameter initialization and online feature extraction provide basic data support.

[0065] S2: Construct a sliding history window with constraints for the prediction time: like Figure 3 As shown, Figure 3 This is a schematic diagram of a sliding window creation process provided in one embodiment of this application. After obtaining the preprocessed historical wind power sequence, this step constructs a sliding historical window according to the prediction time. This sliding historical window is used to limit subsequent... The decomposed data range ensures that each mode decomposition and feature extraction relies only on historical power data before the prediction time, without using data from the prediction target time and beyond, thus guaranteeing that the feature construction process conforms to the time constraints of online wind power prediction.

[0066] Let the preprocessed historical sequence of wind power be... The corresponding sampling time sequence is .in, Indicates the first Wind power values ​​at each sampling time. Indicates the corresponding sampling time. Indicates the sequence length. This indicates the sampling interval between adjacent sampling points.

[0067]

[0068] in, For the first Wind power at each sampling time, For the corresponding sampling time, This represents the total length of the sequence.

[0069] For any prediction anchor point Let the length of the history window be... The predicted step size is The sliding step size is Among them, the historical window length Indicates the number of historical sampling points used in each prediction; prediction step size. The sliding step size represents the number of sampling points the target will move forward relative to the current prediction anchor point; This indicates the movement interval between two adjacent history windows.

[0070] For prediction anchor point i, construct the corresponding sliding history window. This window is composed of prediction anchor points. and the preceding succession It consists of 10 historical sampling points.

[0071]

[0072] The prediction target corresponding to this window is:

[0073] in, For the historical window corresponding to the current prediction anchor point, This represents the future forecast target corresponding to this window. Because... The last sampling point is The prediction target is Therefore, the maximum time of the data contained in the current window is earlier than the target time of the prediction.

[0074] The timing constraints are as follows:

[0075] Therefore, in prediction At that time, window W i Only contains Historical power data prior to this, excluding as well as Subsequent data. Follow-up. Decomposition only The modal decomposition and feature construction are performed within a defined historical window, rather than through a global decomposition of the complete wind power sequence. Therefore, data from the target time and beyond will not participate in the modal decomposition and feature construction process of the current window.

[0076] During continuous prediction, the sliding window follows a preset sliding step size. Move forward along the time axis. The prediction anchor index of a sliding window can be represented as:

[0077] Then the first The sliding history window is as follows:

[0078] The corresponding prediction target is:

[0079] To ensure that both the sliding window and the prediction target exist, the prediction anchor point must satisfy the following:

[0080] Therefore, the number of valid windows that can be generated is:

[0081] In a preferred embodiment, the sliding step size s is 1, meaning the window moves forward by one sampling point each time. At this time, two adjacent sliding windows are respectively:

[0082] From the above relationship, it can be seen that when At that time, two adjacent windows share For each sampling point, only the earliest sampling point moves out of the window, while a new historical sampling point enters the window simultaneously. This high overlap characteristic ensures the continuity of local wind power variation characteristics corresponding to adjacent windows, providing a basis for subsequent... Parameter state inheritance, triggered parameter updates, and modal alignment provide the foundation for data structures.

[0083] The final output of each effective sliding history window is a history power vector. and as a follow-up Input to the decomposition module.

[0084]

[0085] in, This is the historical power input vector corresponding to the current prediction anchor point. (Subsequent...) Decomposition only acts on And generate multi-scale within this window. Quantity. Due to It does not include data for the target time and beyond, therefore it is... Received The features meet the temporal consistency requirements in online prediction.

[0086] Furthermore, to facilitate subsequent window management, parameter state inheritance, and modal alignment, each valid sliding window can be accompanied by output window index information. This window index information describes the window itself and its sequential relationship with adjacent windows. The window index information can be represented as:

[0087] in, Indicates the current window number. Indicates the start time of the window. Indicates the end time of the window. Indicates the predicted target time. Indicates the prediction step size Indicates the sliding step size. Indicates window quality rating. Indicates the number of the previous sliding window.

[0088] It should be noted that, Parameter states are not generated in this step. This step only constructs the sliding history window under the constraint of the prediction time and outputs the historical power vector and window index information. The previous window's... The parameter status will be invoked in subsequent parameter status creation and inheritance steps, based on the window number or window order.

[0089] Through the above methods, this step achieves the construction of historical windows for online forecasting. Firstly, it ensures that the current window only contains data prior to the forecast target, thus preventing future information from interfering with subsequent data. First, it decomposes the data; second, it utilizes the high overlap between consecutive sliding windows to provide a basis for subsequent parameter state inheritance and local updates; third, it preserves the order relationship between windows through window index information to provide a data organization basis for subsequent modal alignment and stable feature output.

[0090] S3: Reconstruct the two-dimensional window matrix and initialize it. Parameter status: Figure 4 This is a schematic diagram of the initial VMD parameter optimization and state caching process provided in one embodiment of this application, as shown below. Figure 4 As shown, after constructing the sliding history window for the prediction time constraint in S2, this step organizes multiple effective sliding history windows into a two-dimensional window matrix in chronological order, and completes the process based on the initial window samples. Initial parameter optimization and parameter state saving. The purpose of this step is to provide a stable initial parameter benchmark for subsequent sliding window decomposition, so that subsequent windows can inherit and update based on the existing parameter state, avoiding the need for each window to perform a global parameter search again.

[0091] S31: Reconstruct the two-dimensional window matrix: From S2, we know that the first The prediction anchor points corresponding to each sliding historical window are: Its window vector is Arrange all valid window vectors row by row according to the chronological order of the prediction anchor points to obtain a two-dimensional window matrix. .

[0092]

[0093] in, Indicates the number of effective sliding windows. This represents the number of historical sampling points for each window. Two-dimensional window matrix. For batch storage and index management of sliding windows only, and will be used later. The decomposition is performed window-by-window, with each row of the matrix as the basic processing object, rather than on a per-matrix basis. It involves a joint decomposition of the entire system, rather than a global decomposition of the complete wind power sequence.

[0094] S32: Determine initial parameters for calibration samples: In order to obtain a stable initial Parameters can be obtained from a two-dimensional window matrix. Before the election A valid window is used as the initial parameter calibration sample group, denoted as . .

[0095]

[0096] in, This indicates the number of windows participating in the initial parameter optimization. If computational requirements are high, it can also be set to... That is, only the first valid sliding window Use multiple initial windows as calibration samples. When calibrating using multiple initial windows, the impact of abnormal fluctuations in a single window on the selection of initial parameters can be reduced.

[0097] S33: Construction Initial parameter candidate set: The main parameters that need to be determined in the decomposition include the number of modes. and penalty factor .in, These are discrete integer parameters used to determine the number of IMFs obtained from the decomposition. Used to control the bandwidth of each mode. Too small a value may cause modal aliasing; if If it is too large, it may generate redundant modes. If... If it is too small, the modal bandwidth may be too wide; if If it is too large, it may weaken the local fluctuation characteristics.

[0098] This invention uses a finite candidate set optimization method to determine the initial parameters, instead of employing a stochastic intelligent optimization algorithm. The reason for this is that... The initial parameters mainly involve low-dimensional, semi-discrete parameters. Using a finite candidate set can ensure that the optimization process is stable, reproducible, and has a controllable computational load.

[0099]

[0100] in, Represents the candidate set of modal numbers. Represents the candidate set of penalty factors. This indicates the number of candidate penalty factors. Candidate penalty factor values ​​can be set at linear or logarithmic intervals.

[0101] The initial parameter candidate combinations are:

[0102] S34: Perform candidate parameter processing Decomposition and scoring: For the candidate set Any combination of parameters The initial parameters were calibrated for the sample groups respectively. Window execution in Decompose to obtain the corresponding Components and center frequency.

[0103]

[0104] in, Indicates the first parameter in the initial parameter calibration sample group. One window, Indicates the first The first window decomposition yields the... indivual Quantity, This indicates the corresponding center frequency.

[0105] To avoid over-decomposition caused by using reconstruction error as the sole evaluation index, this invention constructs a comprehensive scoring function for initial parameters from four aspects: reconstruction accuracy, modal correlation, frequency separation, and modal redundancy.

[0106] First, the first All windows The components are added together to obtain the reconstructed sequence of the window:

[0107] in, Indicates the first Each window is composed of The sequence obtained by component reconstruction.

[0108] Constructing reconstruction error index , used to measure The component's ability to reconstruct the original window sequence:

[0109] in, Represents the original window With reconstruction sequence The root mean square error between them; Display window The standard deviation is used to characterize the fluctuation range of the power data in this window. It is a very small positive number, used to avoid the denominator being zero. The smaller the value, the higher the reconstruction accuracy under that parameter combination.

[0110] Construct modal correlation index Used to measure different Is there strong information redundancy between the components?

[0111] in, Indicates the first The first window indivual Components and the first indivual The correlation coefficient between the components; Indicates selecting two different Components. The smaller this indicator, the more distinct the components. The weaker the repeatability between components, the better the modal independence.

[0112] Constructing frequency separation penalty index Used to measure different The degree of separation of components in the frequency domain. First, separate each... The components are sorted from low to high center frequency, and then the minimum interval between adjacent center frequencies is calculated:

[0113] in, Indicates the first Adjacent windows The minimum interval between center frequencies. If adjacent center frequencies are too close, it indicates that different modes may have frequency band overlap or mode aliasing. Since this indicator uses the reciprocal of the minimum interval, therefore... The smaller the value, the better the center frequency separation effect.

[0114] Constructing a redundancy mode penalty index It is used to measure whether the current combination of parameters produces energy that is too low or has weak physical meaning. Quantity. (Number) The first window indivual The energy percentage of a component is defined as:

[0115] in, Indicates the first The first window indivual The energy percentage of the component. If Less than the preset energy threshold Then it is considered that The component is a low-energy mode.

[0116] The redundancy mode penalty index is defined as:

[0117] in, Indicates the first The energy percentage in each window is below the threshold. of Quantity of components. The smaller the value, the fewer low-energy redundant modes are generated by the combination of parameters.

[0118] To avoid the impact of differences in the dimensions of different evaluation indicators on the overall score, it is possible to... Normalization was performed, and the normalized indicators were denoted as follows: .

[0119] The comprehensive scoring function is defined as follows:

[0120] in, Indicates candidate parameter combinations Overall score; These are the weighting coefficients for each evaluation indicator. All four types of indicators are set according to the principle that "the smaller the value, the better the effect." The smaller the value, the better the overall performance of the candidate parameter combination in terms of reconstruction accuracy, modal independence, frequency separation, and modal effectiveness.

[0121] S35: Determine the initial optimal parameters and complete the initial decomposition: Traverse the set of candidate parameters Calculate the comprehensive score for each set of parameters, and select the parameter combination with the smallest comprehensive score as the initial parameter set. parameter:

[0122] in, This represents the initial optimal number of modes. This represents the initial optimal penalty factor.

[0123] After determining the initial optimal parameters, use them as the first effective sliding window. The decomposition parameters, namely:

[0124] in, and These represent the first effective sliding window. In practice Modal number and penalty factor.

[0125] Use parameters For the first valid sliding window implement Decompose to obtain indivual Components and their corresponding center frequencies:

[0126] After decomposition, the components are arranged according to their center frequencies from low to high. The components are sorted to obtain the first window. sorting relation .

[0127]

[0128] in, Used to record each item in the first valid sliding window The channel order of the components provides an initial reference for subsequent window modal alignment.

[0129] S36: Calculate the initial window Energy percentage and save parameter status: To record each of the initial windows The energy contribution of each component is used to calculate the first effective sliding window. China The percentage of energy in a component.

[0130]

[0131] in, Indicates the first valid sliding window. indivual The energy percentage of the component Indicates the first window indivual Quantity.

[0132] Construct the center frequency vector from the center frequency and energy percentage of the first window. and energy percentage vector .

[0133]

[0134] in, Indicates the first valid sliding window Center frequency vector, Indicates the first valid sliding window Energy percentage vector.

[0135] After completing the initial window decomposition, the first effective sliding window will use... The parameters and their decomposition features are saved as parameter states.

[0136]

[0137] Through the above process, this step completes the determination of the initial optimal parameters and the initial window. The process involves decomposing, saving, and passing the parameter state to subsequent windows. Therefore, subsequent sliding windows do not need to perform global parameter finding again; instead, they continue to decompose based on the parameter state of the previous window, reducing the burden of continuous sliding. Computational complexity and retention Modal continuity provides the foundation.

[0138] Among them, the window number ID is used to identify each window, the modality number K and the penalty factor α are used to provide parameters for subsequent windows to inherit, the center frequency vector F is used to judge modal drift, the energy proportion vector E is used to judge redundant modes, the IMF sorting relationship R is used to perform modal alignment, and the comprehensive score value J is used to evaluate the parameter status.

[0139] S4: Based on the parameter state of the previous window, configure the current window. Initial decomposition: Figure 5 This is a schematic diagram of the initial VMD decomposition process of the current window in the inheritance state provided in one embodiment of this application. After saving the initial window parameter state, for the subsequent... A sliding window ,in This step reads the parameter status of the previous valid window from the parameter status cache table and sets the previous window's... The parameters are used as the initial decomposition parameters for the current window. The purpose of this step is to leverage the high overlap between adjacent sliding windows to avoid re-optimizing global parameters for each window, thereby reducing the impact of continuous sliding. The computational cost of decomposition.

[0140] S41: Read the status of the previous valid window parameters: For the current window Read the parameter status of the previous valid window from the parameter status cache table. If the previous valid window is... Then its parameter state is represented as:

[0141] in, Indicates the parameter status of the previous active window; Indicates the previous window number; This indicates the number of modes used in the previous window; This indicates the penalty factor used in the previous window; Indicates the previous window Center frequency vector; Indicates the previous window Energy percentage vector; Indicates the previous window Sorting order; This indicates the overall score from the previous window.

[0142] In the above parameter states, and Initialize for the current window break down; , and Used for modal drift detection and modal alignment between subsequent windows.

[0143] S42: Inherit from the previous window parameter: Current window The modality number and penalty factor of the previous valid window are inherited first, and used as the initial decomposition parameters for the current window:

[0144] in, Indicates the first The initial number of modalities inherited by each window Indicates the first The initial penalty factor is inherited by each window. (Superscript) This indicates that the parameter is the initial decomposition parameter for the current window.

[0145] S43: Execute the current window using inherited parameters Initial decomposition: Use inheritance parameters ( , For the current window implement Initial decomposition yields the initial state of the current window. Components and their corresponding center frequencies.

[0146]

[0147] in, Indicates the first The first window obtained by decomposing under inheritance parameters An initial Quantity; This indicates that The center frequency corresponding to the component; .

[0148] It should be noted that the decomposition result obtained in this step is the initial decomposition result for the current window. Further evaluation is needed, considering both decomposition quality and modal stability. If the evaluation result meets the requirements, the current window can continue to use the inherited parameters; if the evaluation result does not meet the requirements, subsequent steps will trigger a local parameter update. S43: Organize the initial decomposition result for the current window.

[0149] After the initial decomposition is completed, the components are arranged in ascending order of center frequency. The components are sorted to obtain the initial sorting relationship of the current window. And construct the center frequency vector .

[0150]

[0151] in, Indicates the first The first window An initial The energy percentage of the component; This represents the initial energy percentage vector for the current window.

[0152] This yields the initial decomposition result for the current window:

[0153] in, Indicates the first The initial decomposition results for each window. These results will serve as input for subsequent decomposition quality evaluation, modal drift assessment, and local parameter update assessment.

[0154] Through the above process, the current window can complete its initialization based on the parameter state of the previous window. This process decomposes the parameters without re-executing global parameter optimization. This step provides the basis for subsequent determinations of whether inherited parameters are applicable and whether local updates need to be triggered.

[0155] S5: Initial decomposition evaluation and local update trigger judgment for the current window: Figure 6 This is a schematic diagram of the initial decomposition evaluation and local update triggering judgment process of the current window provided in one embodiment of this application. In S4, the current window... Initialization has been completed based on the parameter state of the previous window. Decompose and obtain the initial parameter state of the current window. This step evaluates the quality and stability of the initial decomposition results of the current window based on the aforementioned comprehensive evaluation index system, and determines whether the current window can continue to use inherited parameters or whether it needs to proceed to the subsequent local parameter update step.

[0156] S51: Calculate the current window's evaluation index: Regarding the initial parameter state of the current window According to the evaluation index system established in S34, the decomposition quality index and modal stability index of the current window are calculated. The indexes include reconstruction error index, modal correlation index, frequency separation penalty index, redundant mode index, and center frequency drift index and energy ratio drift index between adjacent windows.

[0157] The current set of evaluation indicators can be represented as:

[0158] in, This represents the reconstruction error index of the initial decomposition results of the current window; Indicates modal correlation index; Indicates the frequency separation penalty index; Indicates the redundancy modal index; This indicates the center frequency drift index of the current window relative to the previous window; This indicates the energy percentage drift of the current window relative to the previous window.

[0159] in, to These correspond to the reconstruction error, modal correlation, frequency separation penalty, and redundant mode indexes defined in S34, respectively, and their calculation methods are consistent with the aforementioned evaluation index system. To further evaluate the relationship between adjacent windows... To ensure the continuity of modal characteristics, this step further defines the center frequency drift index. and energy percentage drift index .

[0160] The center frequency drift index of the current window relative to the previous window is defined as:

[0161] in, Indicates the first The degree of center frequency drift of each window relative to the previous window; This represents the first decomposition obtained from the initial decomposition of the current window. indivual Component center frequency; Indicates the final saved state of the previous window. The Middle indivual Component center frequency; This represents the initial number of modalities inherited by the current window; This is a very small positive number, used to avoid a denominator of zero. The smaller this index, the more likely adjacent windows have the same number. The smaller the change in the center frequency of a component, the more stable the meaning of the modal frequency.

[0162] The energy percentage drift metric of the current window relative to the previous window is defined as:

[0163] in, Indicates the first The window relative to the previous window The degree of change in energy distribution; This represents the first decomposition obtained from the initial decomposition of the current window. indivual Percentage of energy content; Indicates the final saved state of the previous window. The Middle indivual The percentage of component energy. The smaller this indicator, the more pronounced the energy distribution within adjacent windows. The closer the energy contributions of the components, the more stable the modal energy distribution.

[0164] It should be noted that in this step... and In the current window Sort by center frequency and compared with the previous window The calculation is based on maintaining the correspondence between channels. This is because the current window first inherits the modal count from the previous window. Therefore, in the initial evaluation stage:

[0165] Therefore, the current window and the previous window can be aligned in the same way. The channel number is used to calculate the center frequency drift and energy percentage drift.

[0166] S52: Construct the overall evaluation value for the current window: To comprehensively evaluate the decomposition effect after the current window inherits parameters, the above indicators are normalized. The normalized indicators are denoted as follows: The current window's overall evaluation value is defined as:

[0167] in, Indicates the first The first window Normalized results of initial evaluation indicators for the class; This represents a normalization function used to eliminate differences in dimensions and numerical ranges between different indicators.

[0168] The current window's overall evaluation value is defined as:

[0169] in, Indicates the first The overall evaluation value of each window under inherited parameters; to These represent the six weighting coefficients for the current window's comprehensive evaluation phase, corresponding to six categories of indicators: reconstruction error, modal correlation, frequency separation penalty, redundant modes, center frequency drift, and energy percentage drift.

[0170] The weighting coefficients satisfy:

[0171] in, Indicates the first The weights corresponding to the evaluation indicators. The smaller the value, the better the overall performance of the initial decomposition result of the current window in terms of decomposition quality and modal continuity between windows.

[0172] S53: Two-layer determination of whether the current window uses inherited parameters: To avoid the averaging out of individual abnormal indicators due to relying solely on the comprehensive evaluation value, this invention employs a two-layer judgment mechanism to determine whether the inherited parameters can be used in the current window.

[0173] The first layer is a comprehensive evaluation judgment. If the comprehensive evaluation value of the current window meets the preset comprehensive threshold, it proceeds to the second layer of single-item anomaly check; if it does not meet the threshold, it triggers the subsequent local parameter update step.

[0174]

[0175] in, The threshold is used for comprehensive evaluation.

[0176] The second layer involves checking for individual abnormal indicators. If the overall evaluation meets the requirements, a further check is performed to see if any individual evaluation indicator exceeds the corresponding abnormal limit.

[0177] in, Indicates the first The upper limit of anomalies allowed for each evaluation metric. If any single metric exceeds the corresponding upper limit of anomalies, it is considered that there is a significant decomposition anomaly or modal drift anomaly in the current window, and a local parameter update needs to be triggered.

[0178] S54: Generate a local update trigger flag: Based on the above two-level judgment results, define the partial update trigger flag for the current window. When the overall evaluation value meets the threshold requirement, and all individual anomaly check conditions are met, let:

[0179] At this point, the current window can reuse the inherited parameters and save the initial parameter state of the current window as the final parameter state of the current window:

[0180] When the overall evaluation value does not meet the threshold requirement, or at least one individual evaluation indicator exceeds the abnormal upper limit, let:

[0181] At this point, the initial parameter state of the current window is preserved. and its set of evaluation indicators This is then passed to the subsequent local candidate parameter construction and continuous constraint scoring steps.

[0182] Through the above process, this step determines the validity of the inherited parameters for the current window. This determination mechanism utilizes a comprehensive evaluation value to assess the overall decomposition effect, while also preventing severe anomalies in individual indicators from being masked by the overall score through individual anomaly checks, thereby improving scrolling performance. Stability and reliability of parameter inheritance and triggered update processes.

[0183] S6: Local candidate parameter update and parameter smoothing: Figure 7 This is a schematic diagram of a local candidate parameter update and parameter smoothing process provided in one embodiment of this application. When a local update trigger flag is obtained in S5... When, it indicates the current window The initial decomposition result after directly inheriting the parameters from the previous window does not meet the requirements. This step will not re-traverse the global candidate parameter set, but will instead proceed based on the final parameter state of the previous window. A local candidate parameter set is constructed in the vicinity, and candidate update parameters for the current window are selected through comprehensive scoring; subsequently, the update parameters are smoothed to avoid interference between adjacent windows. Parameters and The modal structure changes frequently.

[0184] S61: Constructing a local candidate parameter set: The number of modalities ultimately saved in the above window and penalty factor Based on this, construct a local candidate parameter set for the current window. The local modality number candidate set is defined as follows:

[0185] in, Indicates the first A candidate set of local modal numbers for each window; This indicates the local search range for the number of modes. The number of candidate modes must satisfy a preset range constraint:

[0186] The candidate set of local penalty factors is defined as follows:

[0187] in, Indicates the first A candidate set of local penalty factors for each window; This represents the preset scaling factor, used to form a finite number of candidate values ​​near the penalty factor of the previous window.

[0188] This yields the current window's local candidate parameter set:

[0189] in, Indicates the first A local candidate parameter set for each window. This set only searches within the vicinity of the parameters of the previous window, avoiding the need to re-optimize global parameters for each sliding window, thus reducing online performance. The computational cost of feature construction.

[0190] S62: Candidate Parameter Scoring and Local Optimal Parameter Selection: For the local candidate parameter set Any combination of candidate parameters is denoted as:

[0191] Use this candidate parameter for the current window implement Decompose the candidate parameter and calculate the set of evaluation indicators corresponding to it according to the evaluation indicator system in S5:

[0192] in, to This represents the decomposition quality index under candidate parameters; and This represents the modal stability index of the current window relative to the previous window under candidate parameters.

[0193] After normalizing the above evaluation indicators, the comprehensive score of the candidate parameters is defined as follows:

[0194] in, Indicates the first A comprehensive score for the group of candidate parameters; Indicates the first The first group of candidate parameters Normalized results of class evaluation indicators; Indicates the first The weights corresponding to the evaluation indicators. The smaller the value, the better the candidate parameter performs in terms of both the quality of the current window decomposition and the modal continuity between windows.

[0195] Select the parameter with the smallest overall score from the local candidate parameter set as the candidate optimal parameter for the current window:

[0196] in, and They represent the first The number of candidate optimal modes and the candidate optimal penalty factor are obtained by searching for local candidate parameters in each window.

[0197] S63: Smooth parameter update and final state saving: To avoid abrupt parameter changes between adjacent windows, this step smooths the candidate optimal parameters. For continuous parameters... Updated using a weighted smoothing method:

[0198] in, Indicates the first The penalty factor ultimately applied to each window; Let represent the smoothing coefficient, and satisfy:

[0199] when When the value is larger, the current window penalty factor is closer to the parameters of the previous window, resulting in a smoother update process; when... When the value is smaller, the current window penalty factor is closer to the candidate optimal value obtained by local search.

[0200] For discrete parameters An update method based on a score improvement threshold is adopted. The modality count is only updated when the candidate optimal parameter shows a significant score improvement relative to the inherited parameter; otherwise, the modality count from the previous window is used.

[0201] in, Indicates the first The final number of modes used by each window; This represents the overall score corresponding to the candidate optimal parameters; This represents the initial overall evaluation value when the current window inherits the parameters from the previous window; This indicates the threshold for score improvement. This method avoids the need to consider modalities. The score fluctuates frequently between adjacent windows due to slight variations.

[0202] Obtain the final parameters and Then, use this parameter on the current window. Execute final break down:

[0203] Further calculate the center frequency vector based on the final decomposition results. Energy percentage vector Sorting Relationship and comprehensive evaluation value And save the final parameter state of the current window:

[0204] in, Indicates the first The final parameter state saved in each window. This state will be passed to the next sliding window for parameter inheritance, modal stability evaluation, and local parameter update judgment in subsequent windows.

[0205] Through the above process, this step constructs a finite set of candidate parameters only near the parameters of the previous window when a local update is triggered, reducing the computational cost of window-by-window global parameter search; at the same time, it suppresses the sum of parameters between adjacent windows through smooth parameter updates. Frequent changes in modal structure make continuous sliding Its features are more suitable for online wind power forecasting tasks.

[0206] S7: Modal Channel Alignment and Stabilized Multiscale Feature matrix output: Figure 8 This is a schematic diagram of modal channel alignment and stable multi-scale IMF feature output provided in one embodiment of this application. After processing S3 to S6, each sliding window has been completed. Decompose and save the corresponding final parameter state. However, during the continuous sliding window decomposition process, even after parameter inheritance, local updates, and smoothing, modality counts may still be affected. Changes, center frequency shifts, or modal rearrangements during abrupt changes, etc. If directly based on... Inputting the original sequence number into the prediction model may result in the same input channel corresponding to different frequency meanings in different windows.

[0207] Therefore, this step sets up a fixed output channel and adjusts the output for different windows according to the center frequency. The components are aligned to channels, resulting in a multi-scale output with fixed dimensions and relatively stable frequency meanings. Feature matrix.

[0208] S71: Set a fixed output channel: To ensure that the input dimensions of subsequent prediction models remain consistent, a fixed number of output channels is first set. In one implementation, the following may be adopted:

[0209] in, This represents the fixed value that is ultimately output to the prediction model. Number of channels; The foregoing The maximum number of modes in the candidate range of mode numbers. By letting... Even if some windows actually decompose to obtain the number of modes Less than Alternatively, zero-padding can be used to maintain consistent input dimensions.

[0210] Furthermore, construct a set of fixed-frequency anchor points:

[0211] in, Represents a set of anchor points with a fixed frequency; Indicates the first The reference center frequency corresponding to each fixed output channel; The frequency anchor point can be determined based on the initial window center frequency vector. The center frequency statistics during the initial parameter calibration phase, or the frequency range division obtained from the preset range.

[0212] By setting a fixed output channel and frequency anchor point, different sliding window decompositions are obtained. The components can be mapped to a uniform frequency scale channel.

[0213] S72: Based on the center frequency Channel mapping: For the The center frequency vector obtained from the decomposition of the window is:

[0214] in, Indicates the first The first window indivual The center frequency of the component; Indicates the first The final number of modalities used by each window.

[0215] To avoid simply following Sequence number alignment causes modal misalignment; this step will resolve this issue for each... Components are mapped to the nearest fixed-frequency anchor channel:

[0216] in, Indicates the first The first window indivual The components are mapped to fixed output channel numbers. This mapping method is based on the center frequency distance, ensuring that components with similar center frequencies are mapped to the correct output channels. Components enter the same fixed channel, thereby improving the connection between different windows. Comparability of feature channels.

[0217] S73: Channel merging, zero padding, and valid markers: Based on the mapping results, define the first... Each fixed output channel corresponds to Number set:

[0218] in, Indicates the first The window is mapped to the first One fixed output channel A set of numbers.

[0219] No. The aligned IMF sequence for a fixed output channel is defined as follows:

[0220] in, Indicates the first After the first window is aligned A fixed aisle; Indicates the first The nth window decomposition yields the nth... Quantity.

[0221] When multiple When components are mapped to the same fixed channel, channel merging is performed through summation to represent the combined components at that frequency scale; when a fixed channel does not have... During component mapping, this channel is treated as an empty channel and filled with zero vectors. This is to distinguish it from true low-energy components. For empty channels, this step sets the channel validity flag.

[0222] First, define the number of channel mappings:

[0223] in, Indicates the first The window is mapped to the first A fixed channel Quantity of components.

[0224] The channel valid marker is defined as:

[0225] in, Indicates the first The window is mapped to the first A valid marker for a fixed channel; Indicates an indicator function. When hour, This indicates that the channel has a corresponding weight; when hour, This indicates that the channel is empty.

[0226] By merging channels and padding empty channels with zeros, the actual number of modes in different windows can be adjusted. Even if they are inconsistent, the number of output channels can be kept constant. .

[0227] S74: Set mutation window flag: For windows with significant abrupt changes in wind power, mode number variations, or center frequency drift, smoothing updates alone may not be sufficient to completely eliminate modal structure changes. To enable subsequent prediction models to identify these specific time periods, this step sets abrupt window marker. .

[0228] First, define the mutation detection variable:

[0229] in, Indicates the first The mutation judgment quantity for each window; and These represent the final modal number of the current window and the previous window, respectively; This indicates the center frequency drift index of the current window relative to the previous window; This indicates the upper limit of the allowable anomalies for the center frequency drift index; It is a very small positive number.

[0230] The mutation window marker is defined as:

[0231] in, Indicates the first The mutation window marker for each window. When When this occurs, it indicates that there is a change in the number of modes or an abnormal drift in the center frequency of the current window; when When the value is set to 0, it indicates that the current window modal structure is relatively stable. This marker can be used as an auxiliary feature input to subsequent prediction models.

[0232] S75: Output Stable Multiscale Feature matrix: After completing modal channel alignment, the first Stable multiscale windows The characteristic matrix is ​​defined as:

[0233] in, Indicates the first Stable multi-scale after window alignment The feature matrix consists of a fixed-frequency scale channel for each row and a time sampling point within a window for each column.

[0234] Simultaneously, the effective channel labeling and mutation window labeling are combined into auxiliary features:

[0235] in, Indicates the first The auxiliary label vector corresponding to each window is used to indicate whether each fixed channel is valid and whether the current window is a mutation window.

[0236] Stabilize all windows The feature matrices are combined in chronological order to obtain the final output:

[0237] in, Represents a stable multi-scale structure consisting of all sliding windows. Feature output.

[0238] Through the above process, this step decomposes different windows to obtain Components are mapped to a fixed output channel, implementing a sliding window. Modal alignment of decomposition results. For periods of mode number variation or abrupt change, input dimension stability is maintained through channel merging, zero padding of empty channels, effective channel labeling, and abrupt window labeling, and modal change information is provided to subsequent prediction models, thereby improving the consistency and identifiability of input features for online wind power prediction.

[0239] The main points of this application are: (1) Sliding under prediction time constraints Feature construction method.

[0240] This invention first defines the sliding window's... The decomposition and feature construction use only historical wind power data prior to the prediction time, excluding data from the future prediction interval. This design aims to avoid data leakage issues in wind power forecasting, ensuring the constructed feature is accurate and robust. The features meet the conditions for real online rolling prediction.

[0241] (2) Continuous based on parameter state inheritance and triggered local update Decomposition mechanism.

[0242] This invention saves and transfers data between sliding windows. The parameter state allows the current window to inherit information such as the number of modes, penalty factor, center frequency, energy percentage, and evaluation results from the previous window. Only when the initial decomposition quality of the current window does not meet the requirements is a local candidate parameter update performed near the parameters of the previous window, and the update results are smoothed. This mechanism avoids completely independent parameter re-finding for each window while maintaining... Decompose the adaptability and continuity of wind power changes.

[0243] (3) Based on center frequency anchor point Channel alignment and fixed-dimensional feature output method.

[0244] This invention sets a fixed output channel and frequency anchor point, maps IMF components obtained from different windows to a unified channel according to the center frequency, and forms a fixed-dimensional multi-scale IMF feature by channel merging and zero-padding of empty channels. This method can solve the problem of different windows Issues such as changes in quantity, misalignment of sequence numbers, and inconsistencies in the meaning of channels enable subsequent prediction models to obtain stable input features.

[0245] Therefore, this application can achieve the following improvements: (1) It can avoid the problem of data leakage in wind power forecasting: This invention explicitly limits each prediction window to using only historical power data prior to the prediction time during the sliding window construction stage, thus enabling... The decomposition and feature construction process conforms to the conditions of real online prediction, avoiding the involvement of future data in feature generation.

[0246] (2) Can improve Decomposing continuity in online sliding prediction: This invention does not decompose each window completely independently. Instead, it uses parameter state caching and parameter inheritance from the previous window to ensure a continuous transmission relationship in the number of modes, penalty factor, center frequency, and energy distribution between adjacent windows, thereby reducing abrupt changes in the decomposition results.

[0247] (3) It can balance parameter adaptability and online computation efficiency: This invention triggers local parameter updates only when the current window decomposition quality does not meet the requirements, and constructs a candidate set only near the parameters of the previous window, avoiding global parameter search for each window, thereby reducing the amount of online computation, while maintaining the ability to adapt to changes in wind power.

[0248] (4) It can output multi-scale data with fixed dimensions and stable channel meanings. feature: This invention utilizes fixed output channels, center frequency anchor point mapping, channel merging, and zero-padding for empty channels to transform the output obtained from different windows. Unify components into a stable channel to resolve issues related to different windows. The issues of quantity variation and channel misalignment facilitate the use of subsequent prediction models.

[0249] In summary, this application can improve performance while meeting online prediction time constraints. The continuity, stability, and engineering applicability of the feature structure.

[0250] This application also provides a feature construction device for online wind power prediction, the device comprising: The acquisition module is used to acquire the historical sequence of wind power of the target wind farm; The processing module is used to construct a sliding historical window corresponding to the prediction time based on the wind power historical sequence according to the window parameters; The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. The processing module is further configured to perform variational mode decomposition on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, corresponding to the prediction time, to obtain multiple target mode components. The processing module is further configured to map each of the target modal components to a fixed output channel according to its corresponding center frequency, thereby generating a fixed-dimensional multi-scale feature.

[0251] This implementation acquires historical wind power sequences and constructs sliding historical windows based on window parameters. Each sliding historical window contains only historical power data up to and including the prediction time, thus limiting the time range upon which feature construction depends from the data source. This ensures that subsequent decomposition and feature generation processes remain consistent with the temporal constraints of online prediction scenarios. When performing variational mode decomposition on the current sliding historical window, the final parameter state of the previous sliding historical window is directly used as the decomposition starting point, eliminating the need for a new global parameter search. This maintains continuity in the number of modes, center frequencies, and energy distribution between adjacent windows, avoiding mode component jumps caused by frequent parameter changes. After decomposition, each target mode component is mapped to a fixed output channel according to its center frequency, forming multi-scale features with fixed dimensions. This ensures that the feature channels output by different sliding historical windows maintain consistency in frequency meaning, thereby providing structurally stable, temporally continuous, and dimensionally uniform input features for subsequent prediction models.

[0252] For other device implementations and technical effects corresponding to the various method implementations, please refer to the relevant methods.

[0253] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0254] Indicatively, such as Figure 9As shown, Figure 9 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 9 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions, such as application programs, that can be executed by the processing component 302. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the methods of any of the embodiments described above.

[0255] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0256] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0257] This application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the method provided in any embodiment.

[0258] Finally, it should be noted that in this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0259] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0260] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A feature construction method for online wind power prediction, characterized in that, The method includes: Obtain the historical wind power sequence of the target wind farm; Based on the window parameters, a sliding historical window corresponding to the prediction time is constructed based on the historical wind power sequence. The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. For the sliding history window at the current moment, based on the final parameter state of the sliding history window at the previous moment, variational mode decomposition corresponding to the prediction moment is performed to obtain multiple target mode components; Each target modal component is mapped to a fixed output channel according to its corresponding center frequency to generate a fixed-dimensional multi-scale feature.

2. The method according to claim 1, characterized in that, The window parameters include the historical window length, prediction step size, and sliding step size; The acquisition of the historical wind power sequence of the target wind farm includes: Obtain the raw wind power data of the target wind farm; The original wind power data is time-aligned and resampled according to a preset sampling interval, and the sampling time corresponding to each sampling point is adjusted to obtain resampled wind power data. The resampled wind power data is validated, abnormal data is removed, and missing data is filled in by interpolation or previous values ​​to form a historical sequence of wind power. And, determine the window parameters.

3. The method according to claim 2, characterized in that, The step of constructing a sliding historical window corresponding to the prediction time based on the wind power historical sequence according to the window parameters includes: Based on the sliding step size, determine the prediction anchor points corresponding to multiple prediction times; Based on the historical window length, the sliding historical window is formed by taking the prediction anchor point and a number of consecutive historical sampling points before it. Wherein, the prediction target corresponding to the sliding history window is the power value of the target sampling point corresponding to the prediction anchor point; The prediction anchor point is moved along the time axis according to the sliding step size to obtain the next sliding history window; Adjacent sliding history windows share multiple sampling points based on the history window length and the sliding step size; Furthermore, the method further includes: Output the window index information corresponding to each of the aforementioned sliding history windows; The window index information is used to indicate the attribute information and order relationship of each sliding history window, including the current window number, window start time, window end time, prediction target time, prediction step size, sliding step size, window quality mark and the previous sliding window number. The window index information is associated with and stored in relation to the historical power vector corresponding to the sliding history window.

4. The method according to any one of claims 1-3, characterized in that, The method further includes: The historical power vectors corresponding to each of the sliding historical windows are arranged according to the order of each of the sliding historical windows to form a two-dimensional window matrix; Based on the two-dimensional window matrix, select the first preset number of sliding history windows as the initial parameter calibration sample group; Construct a candidate set of modality numbers and a candidate set of penalty factors to obtain multiple candidate combinations of window parameters; Wherein, the candidate set of modality numbers is a preset set of discrete integers, and the candidate set of penalty factors is set according to linear intervals or logarithmic intervals; For each sliding history window in the initial parameter calibration sample group, variational mode decomposition is performed under each candidate combination of window parameters to obtain candidate mode components and corresponding center frequencies; A comprehensive score corresponding to the candidate combination of the window parameters is calculated based on at least two of the following indicators: reconstruction accuracy, modal correlation, frequency separation, and modal redundancy. When a single sliding history window is used, the corresponding comprehensive score is directly used as the final score. When multiple sliding history windows are used, the average or median of the comprehensive scores of each sliding history window is used as the final score. The candidate combination of window parameters with the best final score is selected as the initial window parameters. Based on the initial window parameters, the initial variational mode decomposition is performed on the target sliding window to obtain multiple initial component channels. The channel order, center frequency and energy ratio of each initial component channel are determined. The target sliding window includes the first sliding history window in the initial parameter calibration sample group in chronological order; The initial window parameters, the final score, the window number corresponding to the target sliding window, the channel order, the center frequency, and the energy percentage are used as the initial parameter states corresponding to the prediction time.

5. The method according to claim 4, characterized in that, After obtaining multiple target mode components by performing variational mode decomposition corresponding to the prediction time on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, the method further includes: Calculate the initial decomposition quality assessment value corresponding to the current moment; The initial decomposition quality assessment value is obtained by weighted calculation based on multiple preset dimensions. The preset dimensions include the reconstruction accuracy, modal correlation, frequency separation, modal redundancy, center frequency drift and energy ratio drift between the sliding history window at the current time and the sliding history window at the previous time for each target modal component. When the initial decomposition quality assessment value does not meet the preset quality threshold, or when the value of any preset dimension exceeds the preset dimension abnormality threshold, the local parameters of the current sliding history window are updated. When the initial decomposition quality assessment value meets the preset quality threshold, and the values ​​of all preset dimensions do not exceed the preset dimension abnormality threshold, the final parameter state inherited from the previous sliding history window is retained.

6. The method according to claim 5, characterized in that, The step of updating the local parameters of the current sliding history window includes: Based on the neighborhood of the final parameter state of the sliding history window at the previous moment, a local parameter candidate set is constructed. The local parameter combination that optimizes the decomposition quality of the current sliding history window is selected from the local parameter candidate set as the updated parameter state of the current window. Based on the updated parameter state, the initial decomposition parameters inherited by the current sliding history window are locally updated to obtain the updated parameter state. The current sliding history window is then re-executed with variational mode decomposition based on the updated parameter state to obtain the final mode components after the re-execution of variational mode decomposition. The channel order, the center frequency, and the energy percentage are then determined. The updated parameter state, the final score corresponding to the updated parameter state, the window number corresponding to the current sliding history window, the channel order, the center frequency, and the energy percentage are used as the final parameter state corresponding to the current sliding history window.

7. The method according to claim 6, characterized in that, The step of mapping each target modal component to a fixed output channel according to its corresponding center frequency to generate fixed-dimensional multi-scale features includes: Determine the fixed output channel and its corresponding frequency anchor point; For each final modal component obtained by decomposing the current sliding history window, it is mapped to each fixed output channel according to the nearest neighbor relationship or frequency interval affiliation relationship between its center frequency and the frequency anchor point corresponding to each fixed output channel. If multiple final modal components are mapped to the same fixed output channel, then the final modal components are weighted and merged according to energy weights, and the validity flag of the corresponding fixed output channel is set to the first value; If any of the fixed output channels does not have a corresponding final modal component mapping, then the corresponding fixed output channel is padded with zeros, and the validity flag of the corresponding fixed output channel is set to the second value; Based on each of the fixed output channels, determine the fixed-dimensional multi-scale features corresponding to the current sliding history window, and combine them to form a feature matrix corresponding to the current sliding history window; Furthermore, for each of the sliding history windows, based on the center frequency drift, when the difference between the center frequency of the current sliding history window and the center frequency of the corresponding final modal component of the previous sliding history window exceeds a preset drift threshold, and / or when the modal change of the current sliding history window and the corresponding final modal component of the previous sliding history window exceeds a preset modal change threshold, the current sliding history window is determined to be a mutation window, and the validity marker and mutation window marker of each of the fixed output channels in each of the sliding history windows are combined to form an auxiliary feature; Furthermore, the feature matrices corresponding to each of the sliding history windows are combined in chronological order to generate the final output matrix.

8. A feature structure device for online wind power prediction, characterized in that, The device includes: The acquisition module is used to acquire the historical sequence of wind power of the target wind farm; The processing module is used to construct a sliding historical window corresponding to the prediction time based on the wind power historical sequence according to the window parameters; The sliding history window is used to output a historical power vector corresponding to the prediction time. The historical power vector includes multiple data points in the wind power history sequence corresponding to the prediction time and the time before the prediction time. The processing module is further configured to perform variational mode decomposition on the sliding history window at the current time, based on the final parameter state of the sliding history window at the previous time, corresponding to the prediction time, to obtain multiple target mode components. The processing module is further configured to map each of the target modal components to a fixed output channel according to its corresponding center frequency, thereby generating a fixed-dimensional multi-scale feature.

9. A computer device, characterized in that, The method includes one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method as described in any one of claims 1-7.

Images