A Long-Time Series Wind Power Extrapolation Method Based on Multi-Covariate Fusion

By employing a multi-covariate fusion method, dynamic time-delay realignment, adaptive weighting of coupling strength, and error compensation, the problems of time delay and coupling relationship in wind power prediction are solved, thereby improving the accuracy and stability of long-series wind power extrapolation.

CN123069320APending Publication Date: 2026-11-10新疆华电苇湖梁新能源有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611245810.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing wind power forecasting technologies fail to effectively handle the time delay and coupling relationship between meteorological elements and power, resulting in amplified forecasting errors as the extrapolation time history extends, making it difficult to take into account the evolution of both trend and fluctuation components.

Method used

A multi-covariate fusion method is adopted to improve the accuracy and stability of long-series wind power extrapolation by means of dynamic time-delay realignment, adaptive weighting of coupling strength, recursive extrapolation of trend and fluctuation components, and error compensation.

Benefits of technology

It achieves a precise temporal correspondence between changes in meteorological covariates and the impact of power, suppresses interference from low-correlation covariates, takes into account both long-term evolution and short-term random fluctuations, and enhances the accuracy and stability of long-term wind power extrapolation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069320A_ABST
    Figure CN123069320A_ABST
Patent Text Reader

Abstract

This invention discloses a long-term wind power extrapolation method based on multi-covariate fusion, belonging to the field of wind power prediction technology. It includes acquiring multiple meteorological covariates and historical wind power sequences of a wind farm, calculating and realigning the dynamic time-delay offset of each covariate relative to power, measuring the coupling strength between covariates and power within a sliding window, adaptively filtering, and stably weighting and fusing to obtain a covariate fusion representation. Power is decomposed into trend and fluctuation components, and the initial extrapolated power is obtained by recursively superimposing these components based on the fusion representation. Then, an error accumulation factor is estimated and compensated based on the attenuation degree of the representation, and the error accumulation factor is iteratively corrected through extrapolation feedback representation to update the long-term wind power extrapolation result. This invention can improve the temporal accuracy and overall stability of long-term wind power extrapolation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a long-term time-series wind power extrapolation method based on multi-covariate fusion. Background Technology

[0002] Wind power generation, as an important component of clean energy, is significantly influenced by a combination of meteorological factors, including wind speed, wind direction, temperature, air pressure, and humidity, resulting in substantial randomness and volatility in its output power. Long-term wind power extrapolation refers to the process of predicting the output power of wind farms by extending historical meteorological and power data into the future. The prediction results provide crucial support for grid dispatching, turbine operation and maintenance, and electricity trading. The multi-covariate original sequence, composed of multiple meteorological covariate time series and historical wind power series, forms the foundational data for power extrapolation.

[0003] Existing wind power forecasting technologies typically collect meteorological and power data from wind farms directly. These meteorological elements and power data are organized as input features at the same time point, and fixed weights are used to weight the meteorological elements or input them directly into the forecasting model. Then, an autoregressive model or neural network is used to recursively predict the power series, extrapolating the result to the future. Some schemes perform simple smoothing of the power series before forecasting.

[0004] The aforementioned existing technologies, when dealing with the relationship between meteorological elements and power, often ignore the time delay in the impact of meteorological element changes on power. This time delay dynamically changes with the environment, leading to a misalignment of driving information over time. The intensity of the influence of each meteorological element on power varies with time, making it difficult to adapt to the dynamic evolution of the coupling relationship using fixed weights. The evolution patterns of trend components and fluctuation components in the power sequence differ, and an undifferentiated overall recursive approach cannot adequately account for both. As the extrapolation step size increases, the recursive error gradually accumulates, and the power prediction bias amplifies with the extension of the extrapolation time series. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a long-term wind power extrapolation method based on multi-covariate fusion. This method employs dynamic time-delay realignment and adaptive weighted fusion of multiple meteorological covariates, recursively superimposing power trend and fluctuation components, and then using feedback correction and iterative compensation based on the degree of attenuation to improve the temporal accuracy and overall stability of long-term wind power extrapolation.

[0006] The above objectives can be achieved through the following approach: A long-term wind power extrapolation method based on multi-covariate fusion includes obtaining the time series of multiple meteorological covariates and the historical wind power series of a wind farm to obtain the original multi-covariate series; For each covariate in the original multi-covariate sequence, the dynamic time delay offset of each covariate relative to the historical wind power sequence is calculated, and each covariate is realigned along the time axis based on the dynamic time delay offset to obtain the time delay aligned covariate sequence. The coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence is measured within a sliding time window to obtain the covariate coupling strength sequence. Based on the covariate coupling strength sequence, the covariates in the time-delay aligned covariate sequence are weighted and fused to obtain the covariate fusion representation; The historical wind power sequence is decomposed into a power trend component and a power fluctuation component. Driven by covariate fusion representation, the power trend component and the power fluctuation component are progressively extrapolated and superimposed along the extrapolation time history to obtain the initial extrapolated power sequence. The extrapolation error accumulation factor is estimated based on the degree of attenuation of the covariate fusion characterization in the extrapolation time history. The extrapolation error accumulation factor is then used to gradually compensate the initial extrapolated power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

[0007] Based on the same inventive concept, the present invention also provides a long-term wind power extrapolation system based on multi-covariate fusion, the system comprising: The sequence acquisition module is used to acquire the time series of multiple meteorological covariates and the historical wind power series of the wind farm to obtain the original multi-covariate sequence; The time-delay alignment module is used to calculate the dynamic time-delay offset of each covariate relative to the historical wind power sequence for each covariate in the original multi-covariate sequence, and to realign each covariate along the time axis based on the dynamic time-delay offset to obtain the time-delay aligned covariate sequence. The coupling metric module is used to measure the coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence within a sliding time window, and to obtain the covariate coupling strength sequence. The covariate fusion module is used to perform weighted fusion of each covariate in the time-delay aligned covariate sequence based on the covariate coupling strength sequence to obtain the covariate fusion representation. The sequence decomposition module is used to decompose the historical wind power sequence into power trend components and power fluctuation components; The stepwise extrapolation module is used to progressively extrapolate and superimpose the power trend component and the power fluctuation component along the extrapolation time history, driven by the covariate fusion representation, to obtain the initial extrapolated power sequence. The error compensation module is used to estimate the extrapolation error accumulation factor based on the degree of attenuation of the covariate fusion representation in the extrapolation time history, and to use the extrapolation error accumulation factor to gradually compensate the initial extrapolation power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

[0008] Compared with the prior art, the present invention has the following advantages: This invention calculates the dynamic time delay offset of each meteorological covariate and realigns them along the time axis, so that the changes of covariates and their actual impact on power correspond precisely in the time dimension, thereby improving the temporal effectiveness of driving information.

[0009] This invention highlights meteorological components that substantially contribute to power prediction and suppresses interference from low-correlation covariates by measuring the coupling strength between covariates and power within a sliding time window and performing adaptive screening and stable weighted fusion.

[0010] This invention decomposes wind power into trend components and fluctuation components and then recursively superimposes them, taking into account both the long-term evolution of power and short-term random fluctuations, thereby improving the overall fitting ability of long-term extrapolation.

[0011] This invention effectively suppresses the systematic deviation that accumulates as the extrapolation step size increases by estimating the error accumulation factor based on the degree of attenuation and combining it with feedback correction and iterative compensation, thereby enhancing the accuracy and stability of long-term wind power extrapolation. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a long-term wind power extrapolation method based on multi-covariate fusion according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of a long-series wind power extrapolation system based on multi-covariate fusion according to an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Reference Figure 1 One embodiment of the present invention proposes a long-term wind power extrapolation method based on multi-covariate fusion. It adopts dynamic time-delay realignment and adaptive weighted fusion of multiple meteorological covariates, and recursively superimposes the power trend and fluctuation components respectively. Then, it combines feedback correction and iterative compensation of accumulated errors based on the attenuation degree. This method can improve the temporal accuracy and overall stability of long-term wind power extrapolation.

[0016] The method described in this embodiment specifically includes: S1. Obtain the time series of multiple meteorological covariates and the historical wind power series of the wind farm to obtain the original multi-covariate sequence; Specifically, multiple meteorological covariate time series and historical wind power series are acquired from the wind farm's data acquisition system to construct a multi-covariate raw sequence. The meteorological covariate time series refers to the data records of various meteorological elements continuously recorded over time during the wind farm's operation. These meteorological elements include environmental parameters affecting wind power generation, such as wind speed, wind direction, temperature, air pressure, and humidity. The historical wind power series refers to the historical data of the actual electrical power output of the wind farm over time. The multi-covariate raw sequence organizes the multiple meteorological covariate time series and historical wind power series according to their respective sampling times, forming a time series data set containing multi-dimensional meteorological and power information.

[0017] Extracting time range from the wind farm's SCADA system (data acquisition and monitoring system) to Historical data, sampling interval is This yields the discrete time series. The set of meteorological covariates is defined as follows: ,in This represents the total number of meteorological covariates. For the... weather covariates Its time series is denoted as: , in, Indicates the first Each sampling time, , This represents the total length of the time series. Indicates the first The weather covariate in the first The observed values ​​at each moment. The historical wind power series is denoted as: , in, Indicates the wind farm in the 1st The actual output power at time t. The meteorological covariate set is aligned with the historical wind power series to ensure timestamp consistency across series. When missing data is detected at certain times, linear interpolation is used to complete the data. For the t... The weather covariates at time For missing values, interpolation is performed using adjacent valid data points: , in, and These represent the most recent valid observations before and after the missing time point, and These represent the corresponding time points. Outlier detection and handling are performed on each covariate sequence. The calculation of the... Statistical characteristics of a series of meteorological covariates, including the mean. with standard deviation : , , in, Indicates the first The time average of the weather covariates This indicates its standard deviation. When a given observation value... satisfy At that time, it was determined to be an outlier. Median filtering was used to correct the outlier, and the total number of values ​​before and after that time was [data missing]. Replace with the median of the observations at each time point: , in, This represents the half-width of the filter window. This indicates the operation of taking the median of the values ​​within the set. After data cleaning, all meteorological covariate time series and historical wind power series are integrated into a multi-covariate raw sequence matrix. : , Among them, matrix The number of rows is (Corresponding time series length), number of columns is (forward Listed as meteorological covariates, with wind power in the last column.

[0018] For example, a multi-covariate original sequence is constructed based on 300 days of actual operation data from a wind farm. Historical data from the wind farm's SCADA system, spanning from 00:00 on January 1st to 23:45 on October 27th of a certain year, is extracted at sampling intervals. Minutes were used to calculate the total length of the time series. One sampling point. The set of meteorological covariates selected includes wind speed. ,wind direction ,temperature air pressure ,humidity ,common One covariate. Extract wind speed time series. At the first sampling time The observed value corresponding to 00:00 on January 1st of a certain year meters per second; at the second sampling time The observed value corresponding to 00:15 on January 1st of a certain year Meters per second; and so on until the 28800th sampling time. Extract the historical wind power sequence. At the first sampling time Actual output power kilowatts; at the second sampling time Actual output power kilowatts. Temperature sequences were detected during the data preprocessing stage. There is a missing data point at time 1523, which corresponds to 23:30 on January 16th of a certain year. Extract the adjacent valid data points, starting with the temperature at time 1522. Celsius corresponding time Temperature at time 1524 Celsius corresponding time Calculate missing values ​​using the linear interpolation formula. After completing the missing value imputation, the wind speed sequence Perform outlier detection. Calculate the mean of the wind speed sequence. Calculate the standard deviation of the wind speed series. The wind speed observation at time 8456 was detected. meters per second, calculate the deviation meters per second, exceeding Meters per second is considered an outlier. Median filtering is used for correction, with the filter window set to half its width. Five observations were extracted from time points 8454 to 8458: Meters per second, the calculated median is 8.3 meters per second, after correction. meters per second.

[0019] After cleaning all covariates, construct the original sequence matrix of the multi-covariates. The matrix dimension is The first five columns are the time series of wind speed, wind direction, temperature, air pressure, and humidity, respectively, and the sixth column is the historical wind power series, thus obtaining the original multi-covariate series containing complete meteorological and power information.

[0020] S2. For each covariate in the original multi-covariate sequence, calculate the dynamic time delay offset of each covariate relative to the historical wind power sequence, and realign each covariate along the time axis based on the dynamic time delay offset to obtain the time delay aligned covariate sequence. In one embodiment of the present invention, step S2 includes the following steps: Under multiple candidate time delay values, the temporal correlation degree between the shifted subsequence of each covariate and the historical wind power sequence is calculated to obtain the time delay correlation degree distribution; Local extreme values ​​of the time delay correlation distribution are tracked along the time progression direction to obtain the dynamic time delay offset that changes with time. The sampling times of each covariate are shifted and resampled based on the dynamic time delay offset to obtain a time delay aligned covariate sequence.

[0021] Specifically, for each covariate in the original multi-covariate sequence, the dynamic time-lag offset of each covariate relative to the historical wind power sequence is calculated. Based on this dynamic time-lag offset, each covariate is realigned along the time axis to obtain a time-lag aligned covariate sequence. The dynamic time-lag offset refers to the time delay required for changes in meteorological covariates to affect wind power. This delay is not fixed but changes dynamically over time due to factors such as wind field environment and atmospheric stability. The time-lag aligned covariate sequence is a new sequence obtained by shifting and resampling each meteorological covariate along the time axis according to the dynamic time-lag offset, aligning the moment of covariate change with the moment of its actual impact on power in the time dimension.

[0022] Under multiple candidate time delay values, the temporal correlation degree between the shifted subsequences of each covariate and the historical wind power sequence is calculated, yielding the time delay correlation degree distribution. Candidate time delay values ​​refer to preset time delay ranges, typically determined based on the physical characteristics of the wind farm, with a range of [range missing]. ,in This represents the maximum possible number of time lag steps. A shifted subsequence refers to a subsequence obtained by shifting the meteorological covariate time series forward or backward along the time axis by a specified number of time lag steps. The time series correlation degree refers to the degree of correlation between the shifted subsequence and the historical wind power series, quantified using a cross-correlation function. The time lag correlation degree distribution refers to the functional distribution of the time series correlation degree calculated under different candidate time lag values ​​as a function of the time lag. For the... weather covariates Define the candidate time delay set as For candidate time delay values Construct translation subsequences : , in, This means shifting the original time forward. The time after a sampling interval. When When the covariate precedes the power change, The time indicates that the covariate lags behind the power change. (In the sliding window) Internal computation of translation subsequences Historical wind power sequence The cross-correlation coefficient between them is used as the temporal correlation degree. Sliding window The central time is Window length is There are 1 sampling point, and the time range within the window is 1. The formula for calculating the cross-correlation coefficient is: , in, Indicates the first Each covariate at time delay The bottom and the center of the window are Cross-correlation coefficient between time and power series, Indicates the translation subsequence in the window The mean within, Indicates the power sequence in the window The mean within the range. For all candidate time delay values. By performing traversal calculations, the time-delay correlation degree distribution function is obtained. This function describes the time at a fixed point. Correlation strength corresponding to different time delays.

[0023] By tracking local extrema of the time-delay correlation distribution along the time progression direction, the dynamic time-delay offset that varies with time is obtained. Local extrema tracking refers to the tracking of local extrema at the center of each sliding window. From the distribution of time-delay correlation Find the time delay that maximizes the correlation. And it tracks the trajectory of the time delay corresponding to the maximum value over time. The dynamic time delay offset is the time delay function obtained by local extremum tracking. , indicating the first The covariates at time 1 The optimal time delay relative to the power sequence. At the center time of each window. The maximum value of the search time lag correlation distribution: , in, Indicates at time The optimal time delay value that maximizes the cross-correlation coefficient. This represents the independent variable that reaches its maximum value. (As the window center time...) By proceeding along the time axis, the dynamic time delay offset sequence is calculated. To avoid drastic jumps in time delay estimation, the dynamic time delay offset sequence is smoothed using a moving average filter:

[0024] in, This represents the smoothed dynamic time delay offset. This represents the half-width of the moving average window. The sampling times of each covariate are shifted and resampled based on the dynamic time-delay offset to obtain a time-delay aligned covariate sequence. The sampling time shift refers to shifting the sampling time based on the smoothed dynamic time-delay offset. This involves shifting the original sampling time of the covariates forward or backward by the corresponding time step. Resampling refers to obtaining covariate values ​​that are strictly aligned with the power sequence time steps on the shifted time axis using interpolation methods. For the first... A weather covariate, at time [time] The time-delay alignment value is calculated as follows. First, determine the reference time after translation: , in, Indicates the covariate at time t. The original timeline position to refer to. Indicates the sampling interval. Because... This may not correspond to the original sampling points, requiring linear interpolation between adjacent sampling points. Assume... Located at the original sampling time and Between, among satisfy Then the covariate value after time delay alignment is: , in, Indicates the first The covariates at time 1 The value after time delay alignment and These represent the original sequence at time [time]. and The observed values. For all times. ( Perform the above translation and resampling operations to obtain the first... A time-delay aligned sequence of covariates: , For all Each meteorological covariate is subjected to time-delay alignment, resulting in a set of time-delay aligned covariate sequences. .

[0025] For example, regarding the wind speed covariate constructed in step S1 Calculate its dynamic time-delay offset relative to the historical wind power sequence. Based on the physical characteristics of the wind farm, set the maximum possible number of time-delay steps. Each sampling interval corresponds to a time delay of [number] times. Minutes, or 5 hours. Candidate time delay set. It contains 41 candidate time delay values. The sliding window length is set to... There are 10 sampling points, corresponding to a time span of 10 ... Minutes, approximately 24 hours. Select the center time of the window. For a given date of January 6th, 04:15, the time range within the window is: For candidate time delay values Constructing a translational subsequence of wind speed This subsequence shifts the original wind speed sequence forward by 8 sampling intervals, or 2 hours. The values ​​of the shifted subsequence and the power sequence are extracted within the window, and the mean of the shifted subsequence within the window is calculated. The translated wind speed values ​​corresponding to times 452 to 548 within the window are added together to calculate the result. Meters per second. Calculate the mean of the power sequence within the window. The numerator of the cross-correlation coefficient is calculated as follows: , Calculate the covariance contribution at each time step. time, meters per second kilowatts, contribution value ;exist time, meters per second, kilowatts, contribution value The contributions from all 97 time points within the window are summed to obtain the numerator 12450. The denominator of the cross-correlation coefficient is then calculated, first by calculating the standard deviation of the shifted subsequences. Calculate the standard deviation of the power series. Calculate time delay Corresponding cross-correlation coefficients For all candidate time delay values Repeat the above calculations. Time calculation ;exist Time calculation ;exist The time has come ;exist Time calculation After iterating through all candidate time lags, it was found that... The corresponding cross-correlation coefficient is the largest, therefore it is determined at time [time]. The optimal time delay is This indicates that wind speed changes precede power changes by 7 sampling intervals, or 105 minutes.

[0026] Move the center of the window along the direction of time progression to Repeat the above calculation to obtain Move to get Move to get The dynamic time-delay offset sequence is smoothed by setting a moving average window width of half. ,calculate Smoothing time delay: The smoothed dynamic time delay offset was obtained as 7.14 sampling intervals. The wind speed covariates were then realigned based on the smoothed time delay offset. At time... Calculate the reference time after translation .time Corresponding to 04:30 on January 6th of a certain year, shifted forward by 107.1 minutes, the reference time is... This corresponds to 02:42.9 on January 6th of a certain year. This reference time is located at the original sampling time. (Corresponding to 02:30) and (Between 02:45 and 02:45). Extract the observations from the original wind speed sequence at these two times. meters per second meters per second, performing linear interpolation The wind speed covariance at time t is obtained. The time-delay aligned value is 7.972 meters per second. Translation and resampling operations were performed on all 28,800 time points to obtain the time-delay aligned covariate sequence of wind speed. This sequence allows for precise alignment of wind speed changes with their actual impact on power over time.

[0027] S3. Measure the coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence within the sliding time window to obtain the covariate coupling strength sequence. In one embodiment of the present invention, step S3 includes the following steps: Within a sliding time window, the fluctuation trends of the time-delay aligned covariate sequence and the historical wind power sequence are extracted to obtain the covariate fluctuation trend and the power fluctuation trend. The consistency of the covariate fluctuation trend and the power fluctuation trend within the same sliding time window is calculated to obtain the in-window consistency measure. The consistency metric within the sliding time window is accumulated and normalized along the advancing direction of the sliding time window to obtain the covariate coupling strength sequence.

[0028] Specifically, within a sliding time window, the coupling strength between each covariate in the time-lag aligned covariate sequence and the historical wind power sequence is measured, resulting in a covariate coupling strength sequence. Coupling strength refers to the degree of interaction between meteorological covariates and wind power. This strength is not a static fixed value but changes dynamically over time, reflecting the actual driving force of the covariates on power at different times. The covariate coupling strength sequence is a sequence formed by arranging the coupling strengths calculated at each time point in chronological order, used for subsequent time-varying weighted fusion of the covariates.

[0029] Within a sliding time window, the fluctuation trends of the time-delay aligned covariate sequence and the historical wind power sequence are extracted separately to obtain the covariate fluctuation trend and the power fluctuation trend. The fluctuation trend refers to the direction and amplitude characteristics of the time series changes between adjacent moments, obtained by calculating the first difference of the sequence. A sliding time window is a fixed-length window that moves along the time axis, containing multiple consecutive sampling moments. The covariate fluctuation trend refers to the rate of change sequence of the time-delay aligned covariate sequence within the sliding window, while the power fluctuation trend refers to the rate of change sequence of the historical wind power sequence within the same window.

[0030] For the A time-delay aligned covariate sequence In the sliding window The fluctuation trend is calculated internally. Window The central time is Window length is There are 1 sampling point, and the time range within the window is 1. The tendency of covariate fluctuations is calculated using first-order differences: , in, Indicates the first The covariates at time 1 The fluctuation trend value, that is, the change relative to the previous moment. Within the window... Internal extraction of covariate fluctuation trend sequence: , The length of the fluctuation trend sequence within the window is Each element (the sequence length is reduced by 1 due to the first-order difference). Similarly, the historical wind power sequence is calculated within a window. Internal power fluctuation trend: , in, Indicates the power sequence at time 10:00 The fluctuation trend value. Within the window. Internal extraction of power fluctuation trend sequence: , The directional consistency between the covariate fluctuation trend and the power fluctuation trend within the same sliding time window is calculated to obtain an intra-window consistency measure. Directional consistency refers to the degree of synchronization between two fluctuation trend sequences in the direction of change; high directional consistency is indicated when the power increases (or decreases) in tandem with the covariate. The intra-window consistency measure is quantified by calculating the sign matching rate of the fluctuation trends. A sign function is defined. For any real number : , In the window Within this context, we compare the signs of the covariate fluctuation trends with those of the power fluctuation trends at each time step, defining a consistency indicator function: , in, Indicates at time An indicator variable for whether two fluctuation trends have the same sign takes a value of 1 when both signs are the same (both positive or both negative), and a value of 0 otherwise. Calculation window The average of the consistency indicator functions over all time points within the window yields the in-window consistency metric. , in, Indicates the first The covariates at the center of the window The in-window consistency measure, with a value range of [value range missing]. A value closer to 1 indicates higher directional consistency. The consistency metric within the sliding time window is accumulated and normalized along the advancing direction to obtain a sequence of covariate coupling strengths. The accumulation operation involves weighted summation of the consistency metrics from historical windows, reflecting the long-term coupling relationship between covariates and power. The normalization operation maps the accumulated values ​​to a uniform scale range, facilitating comparisons between different covariates. A cumulative decay weighting function is defined to assign a weight that decreases over time to the historical consistency metrics: , in, Indicates the time elapsed since the current time The weights corresponding to the historical consistency metrics for each sampling interval This represents the decay time constant, controlling the degree to which historical information is retained. At time... Calculate the cumulative consistency metric: , in, Indicates the first The covariates at time 1 The cumulative consistency measure, with the summation range covering from the start of the sequence. Up to the current moment All previous historical windows. Normalizing the cumulative consistency measure yields the covariate coupling strength sequence: , in, Indicates the first The covariates at time 1 The coupling strength, the denominator is all The sum of the cumulative consistency measures of the covariates, after normalization, equals the sum of the coupling strengths of all covariates, and its value ranges from 1 to 1. For all moments ( ) and all Perform the above calculations on each covariate to obtain a set of covariate coupling strength sequences. .

[0031] For example, the wind speed time-delay aligned covariate sequence obtained in step S2 This measures the coupling strength between the wind power and historical wind power sequences. The sliding window length is set to... There are 10 sampling points, corresponding to a time span of 10 ... Minutes. Select the center time of the window. For a specific date, January 7th, 05:00, the time range within the window is: Calculate the fluctuation trend of the wind speed time-delay aligned covariate within the window. At time... Wind speed alignment value meters per second, the previous moment meters per second, fluctuation trend meters per second, with a positive sign; at time Wind speed alignment value meters per second, fluctuation trend Meters per second, with a positive sign.

[0032] Calculate the fluctuation trend of the power sequence within the same window. At time... ,power kilowatts, a moment ago kilowatts, fluctuation trend Kilowatt, with a positive sign; at time ,power kilowatts, fluctuation trend Kilowatts, with a positive sign. Sign consistency is compared time-by-time. At time... The covariate fluctuation trend is positive, the power fluctuation trend is positive, and the product... Consistency indicator function At any moment Both are positive signs. Continue comparing the various times within the window, at time... Wind speed fluctuations tend to be The sign for meters per second is negative, and the power fluctuation trend is as follows: The kilowatt symbol is positive, and the product is positive. Consistency indicator function .

[0033] The consistency indicator function values ​​were statistically analyzed at 24 fluctuation trend points within the window, with 19 points showing consistent signs and 5 points showing inconsistent signs. The in-window consistency measure was then calculated. The wind speed covariance at the center of the window was obtained. The in-window consistency metric is 0.792.

[0034] Accumulate the consistency metric within the window along the time progression direction. Set a decay time constant. Each sampling interval corresponds to 24 hours. Calculate the weights of historical metrics of different durations from the current time. For a historical metric with one sampling interval from the current time, the weight is... For historical metrics 48 sampling intervals from the current time, the weights are... .Will The cumulative consistency measure of the wind speed covariate is calculated by summing the consistency measures of all previous historical windows according to the attenuation weight. .

[0035] Similarly, calculate the remaining four covariates at time t. The cumulative consistency measure yields the wind direction covariate. Temperature covariate air pressure covariance Humidity covariate The sum of the cumulative consistency measures of the five covariates is The cumulative consistency measure of wind speed covariates is normalized to obtain the coupling strength. The wind speed covariance at time t is obtained. The coupling strength is 0.319. The same calculation is performed on all time points and all covariates to obtain a set of covariate coupling strength sequences, reflecting the dynamic driving force of each meteorological covariate on wind power at different time periods.

[0036] S4. Based on the covariate coupling strength sequence, the covariates in the time-delay aligned covariate sequence are weighted and fused to obtain the covariate fusion representation. In one embodiment of the present invention, step S4 includes the following steps: Based on the adaptive limit determined by the intensity distribution of the covariate coupling strength sequence, covariates with coupling strength below the adaptive limit in the time-delay aligned covariate sequence are removed to obtain the effective covariate set. Based on the covariate coupling strength sequence, time-varying fusion weights are assigned to each covariate in the effective covariate set, thus obtaining the time-varying fusion weights; Applying a smoothing constraint to the time-varying fusion weights along the time direction yields smoothed fusion weights; Calculate the magnitude of change in the smoothed fusion weights between adjacent time points to obtain the weight abrupt change degree; When the weight abruptness exceeds the allowable range determined by the distribution of weight abruptness changes, the smooth fusion weights are pulled back to suppress them, resulting in stable fusion weights. By superimposing the effective covariate set according to the stable fusion weights at each time step, a covariate fusion representation is obtained.

[0037] Specifically, based on the covariate coupling strength sequence, the covariates in the time-delay aligned covariate sequence are weighted and fused to obtain the covariate fusion representation. The covariate fusion representation refers to the comprehensive feature sequence formed by weighting and superimposing multiple meteorological covariates according to their coupling strength with power. This sequence encapsulates the most valuable components for power prediction from multidimensional meteorological information.

[0038] Based on an adaptive threshold determined by the intensity distribution of the covariate coupling strength sequence, covariates with coupling strengths below the adaptive threshold in the time-delay aligned covariate sequence are removed, resulting in the effective covariate set. The adaptive threshold is a dynamically determined selection threshold based on the statistical distribution of the coupling strengths of all covariates at the current time, rather than a fixed constant. The effective covariate set refers to the set of covariates whose coupling strengths reach or exceed the adaptive threshold; these covariates contribute substantially to power prediction. At time... Calculate the mean and standard deviation of the coupling strength of all covariates to determine the adaptive limits: , in, Indicates time The adaptive limit, This represents the mean of the coupling strength of all covariates. The standard deviation of the coupling strength is represented by the standard deviation of the coupling strength. This represents the threshold adjustment coefficient, used to control the stringency of the screening. The mean and standard deviation are calculated as follows: , , For the A covariate, when its coupling strength When this is done, include it in the effective covariate set. Otherwise, they are discarded. The set of effective covariates is represented as: , Based on the covariate coupling strength sequence, time-varying fusion weights are assigned to each covariate in the effective covariate set, resulting in time-varying fusion weights. These time-varying fusion weights are weighting coefficients determined by the coupling strength of the covariates at the current time point and dynamically adjusted over time. The coupling strengths of each covariate in the effective covariate set are then re-normalized to obtain the time-varying fusion weights. , in, Indicates the first The covariates at time 1 The time-varying fusion weights are calculated by dividing the denominator by the sum of the coupling strengths of all covariates within the effective covariate set. After normalization, the sum of the weights within the effective covariate set is 1. A smoothing constraint is applied to the time-varying fusion weights to obtain smoothed fusion weights. The smoothing constraint restricts drastic changes in the fusion weights between adjacent time points, making the evolution of the weights more continuous and stable over time. An exponentially weighted moving average is used to smooth the time-varying fusion weights. , in, Indicates the first The covariates at time 1 Smooth fusion weights, This represents the smoothing factor, with a value range of 1. , This represents the smoothed fusion weights at the previous time step. The magnitude of change in the smoothed fusion weights between adjacent time steps is calculated to obtain the weight abruptness. The weight abruptness refers to the absolute change in the smoothed fusion weights between two consecutive time steps, used to monitor whether abnormal weight jumps occur. , in, Indicates the first The covariates at time 1 The weight abruptness is calculated. When the weight abruptness exceeds the allowable range determined by the distribution of weight abruptness changes, the smoothed fusion weights are pulled back to suppress abnormal jumps, resulting in stable fusion weights. The allowable range refers to the normal fluctuation interval determined based on the historical statistical distribution of weight abruptness. Pull-back suppression means that when the weight abruptness exceeds the allowable range, the current weight is pulled back in the direction of the previous weight, suppressing abnormal jumps. The historical mean of weight abruptness is calculated. with standard deviation Determine the upper limit of the allowable range: , in, This represents the upper bound allowed for the degree of weight mutation. This represents the allowable range adjustment coefficient. When... At that time, the smoothing fusion weights are pulled back to suppress: , in, Indicates the first The covariates at time 1 Stable fusion weights This represents the pullback suppression coefficient, with a value range of [value missing]. The smaller the value, the stronger the pullback. In this case, the stable fusion weights are directly taken as the smooth fusion weights, i.e. The stable fusion weights after pullback suppression are renormalized to ensure that the sum of weights within the effective covariate set is 1. , By superimposing the effective covariate set according to the stable fusion weights at each time step, we obtain the covariate fusion representation: , in, Indicates time The covariate fusion representation is obtained by weighting and summing each covariate in the effective covariate set according to stable fusion weights.

[0039] Because different covariates have different dimensions (e.g., wind speed in meters per second, temperature in degrees Celsius, and air pressure in hectopascals), the time-delay aligned covariate sequences need to be standardized before weighted superposition to eliminate dimensional differences. z-score standardization is used. , in, This represents the standardized time-delay aligned covariate value. and The first The mean and standard deviation of each covariate. After standardization, each covariate is dimensionless; the weighted summation of the covariates represents the overall covariate composition. This is a dimensionless comprehensive characteristic. Therefore, the actual calculation of the covariate fusion representation uses standardized covariates: , For all moments ( Performing the above weighted fusion operation yields the covariate fusion representation sequence. .

[0040] For example, for the covariate coupling strength sequence obtained in step S3, at time... A weighted fusion of five meteorological covariates was performed. The five covariates are known at time [time value missing]. The coupling strengths are respectively wind speed ,wind direction ,temperature air pressure ,humidity Calculate the mean of the coupling strength of the five covariates. Calculate the standard deviation of the coupling strength. Set the limit adjustment coefficient. Calculate the adaptive limit The coupling strength of each covariate is compared with the adaptive limit. Wind speed Inclusion, wind direction Rejection, temperature Intake, air pressure Incorporation, humidity Eliminate covariates to obtain the effective set of covariates. It includes three covariates: wind speed, temperature, and air pressure. The coupling strength within the effective covariate set is renormalized, and the sum of the effective covariate coupling strengths is... Calculate the time-varying fusion weights for wind speed: , Similarly, the time-varying fusion weights for temperature are calculated. Time-varying fusion weights of air pressure Apply smoothing constraints to the time-varying fusion weights and set a smoothing factor. Given the smoothed fusion weights of the wind speed at the previous moment. Calculate the smoothing fusion weights for the current wind speed: Calculate the wind speed weighted abrupt change degree. The historical average of the wind speed weighted abrupt change is known. with standard deviation Set the allowable range adjustment coefficient Calculate the upper limit of tolerance ,because If the weight mutation rate does not exceed the allowable range, the stable fusion weights are directly taken as the smooth fusion weights. The stable fusion weights of the three effective covariates are renormalized. Assume the stable fusion weights for temperature and air pressure are respectively... , The sum of the three is Normalized wind speed weights Similarly, the temperature normalized weights are obtained. Normalized weight of air pressure z-score standardization was performed on the three effective covariates. The wind speed at time [time missing] is known. alignment value meters per second, average wind speed meters per second, standard deviation meters per second, calculate the standardized value Similarly, calculate the standardized temperature value. Standardized air pressure value The standardized effective covariates are weighted and summed according to stable fusion weights to obtain the covariate fusion representation. Get the time The covariate fusion representation is 0.173. Performing the same calculation at all time points yields the covariate fusion representation sequence. This sequence serves as the core driving information for subsequent power extrapolation.

[0041] S5. Decompose the historical wind power sequence into power trend components and power fluctuation components; In one embodiment of the present invention, step S5 includes the following steps: Multi-timescale smoothing was performed on the historical wind power series to obtain smoothed series at different time scales; The power trend component is obtained by extracting the gradient component from smooth sequences at different time scales. The power fluctuation component is obtained by subtracting the historical wind power sequence from the power trend component over time.

[0042] Specifically, the historical wind power series is decomposed into a power trend component and a power fluctuation component. The power trend component refers to the main component of the wind power series that changes slowly over time, reflecting the long-term evolution of power. The power fluctuation component refers to the residual component of the wind power series that fluctuates rapidly around the trend component, reflecting short-term random changes in power.

[0043] Multi-timescale smoothing is applied to historical wind power series to obtain smoothed series at different time scales. Multi-timescale smoothing refers to filtering the power series using multiple smoothing windows of different lengths, with each window length corresponding to a different time scale. The smoothed series is the power series that has been filtered by moving average to eliminate short-term fluctuations and retain gradual variation components.

[0044] Set smoothing scale set ,in Indicates the first The length of a smooth window, This represents the total number of smoothing scales. For the th... A smoothing scale is used to smooth the wind power series using a moving average: , in, Indicates the first Time at a smooth scale The smoothed power value, This represents the length of the smoothing window. Gradual components are extracted from smoothed sequences at different time scales to obtain power trend components. Gradual components refer to low-frequency stable components commonly present in smoothed sequences across multiple time scales. Weighted fusion of smoothed sequences at multiple smoothing scales highlights the gradual components and suppresses differences between different scales. , in, Indicates time The power trend component, Indicates the first The fusion coefficients at a smooth scale satisfy... The fusion coefficient typically assigns higher weights to larger smoothing scales to enhance the stability of the trend. The power fluctuation component is obtained by subtracting the historical wind power series from the power trend component over time. , in, Indicates time The power fluctuation component. The original wind power sequence can be reconstructed by superimposing the power trend component and the power fluctuation component at each moment, satisfying... This ensures the completeness of the decomposition.

[0045] For example, regarding the historical wind power sequence constructed in step S1 Perform trend and volatility decomposition. Define the smoothing scale set. ,Right now Several smoothing scales, corresponding to 1-hour, 4-hour, and 24-hour time scales respectively. Select time point. This corresponds to 05:00 on January 7th of a certain year. On the first smoothing scale... Calculate the smoothed power value and extract the time. to There are a total of 4 power values. kilowatt, kilowatt, kilowatt, kilowatts, calculate smoothing value In the second smoothing scale Below, extraction time to Averaging of 16 power values ​​yielded the following result. kilowatts. At the third smoothing scale. Below, extraction time to A total of 96 power values ​​were averaged to obtain the following result. kilowatts. Set the fusion coefficient. , , The sum of the three factors must be 1, and higher weights are assigned to larger smoothing scales. Calculate the power trend component. Get the time The power trend component is 4744 kW. The power fluctuation component is calculated by subtracting the original power value from the trend component. Get the time The power fluctuation component is 146 kW. Verifying the completeness of the decomposition, the sum of the trend component and the fluctuation component is... The power value in kilowatts is consistent with the original power value, satisfying the completeness requirement of the decomposition. Performing the same calculation at all times yields the power trend component sequence. With power fluctuation component sequence .

[0046] S6. Driven by covariate fusion representation, the power trend component and the power fluctuation component are progressively extrapolated and superimposed along the extrapolation time history to obtain the initial extrapolated power sequence. In one embodiment of the present invention, step S6 includes the following steps: Driven by covariate fusion characterization, the power trend component is progressively extrapolated along the extrapolation time history to obtain the trend extrapolation component. Driven by covariate fusion characterization, the power fluctuation component is progressively extrapolated along the extrapolated time history to obtain the fluctuation extrapolation component. The trend extrapolation component and the fluctuation extrapolation component are aligned and superimposed according to the extrapolation time to obtain the initial extrapolation power sequence.

[0047] Specifically, driven by covariate fusion representation, the power trend component and power fluctuation component are progressively extrapolated and superimposed along the extrapolation time history to obtain the initial extrapolated power sequence. The extrapolation time history refers to the prediction time range extending from the end of the historical data to the future. Progressive extrapolation refers to the process of using known historical values ​​and covariate fusion representation as inputs to extrapolate future power values ​​moment by moment. The initial extrapolated power sequence is the original prediction sequence obtained by superimposing the trend extrapolation component and the fluctuation extrapolation component without error compensation.

[0048] Driven by covariate fusion representation, the power trend component is progressively extrapolated along the extrapolated time history to obtain the trend extrapolated component. The trend extrapolated component refers to the sequence obtained by extending the power trend component into the future. Let the last time of the historical data be denoted as . Extrapolated timeline for the future Each sampling interval, the extrapolated time is denoted as . .

[0049] The recursion of the power trend components adopts an autoregressive structure, combined with covariate fusion representation as an external driver. For the extrapolation time... ( The recursive formula for the trend extrapolation component is: , in, Indicates the extrapolation time The trend extrapolation component, Indicates the first Autoregressive coefficients of order 1 Indicates the order of autoregression. This represents the coefficient of covariate fusion that characterizes the driving force of the trend. This represents the fusion representation of covariates at the extrapolation time. When hour, Take the actual value of the historical power trend component Autoregression coefficient The historical power trend component sequence is determined by least-squares fitting. An autoregressive equation system for the historical trend component is constructed, minimizing the sum of squared residuals: , Driving coefficient The influence of the covariate fusion representation on trend changes is determined by the regression relationship between historical power trend components and the covariate fusion representation. Since the covariate fusion representation is dimensionless, while the power trend component is measured in kilowatts, the driving coefficient... The dimension of kilowatt is kilowatt, which makes It has the dimension of power, which is consistent with the dimension of the autoregressive term.

[0050] Driven by covariate fusion representation, the power fluctuation component is progressively extrapolated along the extrapolated time history to obtain the extrapolated fluctuation component. The extrapolated fluctuation component refers to the sequence obtained by extending the power fluctuation component into the future. The power fluctuation component reflects short-term random fluctuations, and its extrapolation is also driven by an autoregressive structure combined with covariate fusion representation: , in, Indicates the extrapolation time extrapolated components of fluctuations, The first component representing the fluctuation component Autoregressive coefficients of order 1 This indicates the autoregressive order of the fluctuation component. This represents the coefficient of covariate fusion that characterizes the driving force of the change on the fluctuation. This represents the change in the covariate fusion representation between adjacent extrapolated time points. The power fluctuation component is sensitive to short-term changes in the covariates; therefore, the change represented by the covariate fusion is used as the driving factor, and the driving coefficient is... The units are in kilowatts, ensuring that the driving and autoregressive terms have consistent units. The trend extrapolation components and fluctuation extrapolation components are aligned and superimposed according to the extrapolation time to obtain the initial extrapolated power sequence: , in, Indicates the extrapolation time The initial extrapolated power value. For all extrapolated times... By performing recursion and superposition, the initial extrapolated power sequence is obtained. Due to the covariate fusion representation at the extrapolation time. Since direct observation is not possible, extrapolation prediction is performed using the historical evolution patterns of covariate fusion representations. An autoregressive recursion is also established for the covariate fusion representation sequence to obtain the estimated values ​​of the covariate fusion representations over the extrapolated time history, providing the driving input for power recursion.

[0051] For example, an extrapolation recursion is performed on the covariate fusion representation obtained in step S4 and the power trend component and power fluctuation component obtained in step S5. The end time of the historical data is set. Extrapolated timeline for the future Each sampling interval corresponds to 24 hours. The autoregressive order of the trend component is set. The order of the autoregressive component of the fluctuation component Least squares fitting is performed on the historical power trend component series to obtain the autoregressive coefficients of the trend components. , Determining trend-driving coefficients through historical regression. kilowatts. For the first extrapolated time... Perform trend recursion. The last values ​​of the historical trend components are known. kilowatt, kilowatts, covariate fusion representation estimates at extrapolated time Calculate the trend extrapolation component. The trend extrapolation component at the first extrapolation moment is 4739.1 kW.

[0052] Least squares fitting is performed on the historical power fluctuation component sequence to obtain the autoregressive coefficients of the fluctuation components. , , The volatility driving coefficient was determined through historical regression. kilowatts. For the first extrapolated time... Perform fluctuation recursion. The final values ​​of historical fluctuation components are known. kilowatt, kilowatt, Kilowatts, covariate fusion characterizes the amount of change Calculate the extrapolated components of the fluctuation. The extrapolated fluctuation component at the first extrapolated time was obtained as 80.06 kW. The extrapolated trend component and the extrapolated fluctuation component were superimposed to obtain the initial extrapolated power value at the first extrapolated time. The initial extrapolated power at the first extrapolated time step is 4819.16 kW. Continuing the extrapolation along the time path, the power at the second extrapolated time step is... ,Will The trend extrapolation component is used as a known input and substituted into the recursive formula, and the above calculation is repeated. Recursion and superposition are performed for all 96 extrapolation time points to obtain the initial extrapolated power sequence. .

[0053] S7. Estimate the extrapolation error accumulation factor based on the degree of attenuation of the covariate fusion representation in the extrapolation time history, and use the extrapolation error accumulation factor to gradually compensate the initial extrapolation power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

[0054] In one embodiment of the present invention, step S7 includes the following steps: Based on the change in the intensity of the characterization along the extrapolated time history using the covariate fusion characterization, the characterization decay curve is obtained; Based on the degree to which the decay curve decreases with the extrapolation step size, the extrapolation error accumulation factor is obtained; By using the extrapolation error accumulation factor to gradually compensate the initial extrapolated power sequence along the extrapolation time, the long-term wind power extrapolation result is obtained.

[0055] Specifically, the extrapolation error accumulation factor is estimated based on the degree of attenuation of the covariate fusion representation over the extrapolation time series. This accumulation factor is then used to progressively compensate the initial extrapolated power sequence along the extrapolation time, yielding the long-term wind power extrapolation result. The degree of representation attenuation refers to the decrease in representation strength caused by increased information uncertainty during the extrapolation process. The extrapolation error accumulation factor is a correction coefficient used to quantify the prediction error accumulated with increasing extrapolation step size. The long-term wind power extrapolation result refers to the final wind power prediction sequence obtained after error compensation.

[0056] Based on the change in representation intensity of the covariate fusion representation along the extrapolation time history, a representation decay curve is obtained. Representation intensity refers to the amount of effective information carried by the covariate fusion representation at a certain extrapolation time, and is measured by the reliability of the covariate fusion representation estimate. The representation decay curve is a function curve showing how the representation intensity changes with the extrapolation step size.

[0057] Since the covariate fusion representation is obtained through autoregressive recursion over the extrapolated time history, the recursion error gradually accumulates as the extrapolation step size increases, and the reliability of the representation estimate gradually decreases. Define the extrapolation time. Characterization intensity: , in, Indicates the extrapolation time The characterization intensity, with a value range of , Indicates the extrapolation step size. This represents the decay rate coefficient, controlling how quickly the characterization intensity decreases with the extrapolation step size. When The characterization intensity is 1 at time, and it increases with time. Increasing the characterization intensity monotonically decreases, forming a characterization decay curve. Based on the degree to which the characterization decay curve decreases with the extrapolation step size, the extrapolation error accumulation factor is obtained. The extrapolation error accumulation factor reflects the degree of prediction error amplification corresponding to the decrease in characterization intensity; the lower the characterization intensity, the more significant the prediction error accumulation. The extrapolation error accumulation factor is defined as follows: , in, Indicates the extrapolation time Extrapolation error accumulation factor, This represents the error accumulation adjustment coefficient. This represents the decrease in intensity relative to the initial full intensity. When the intensity of the characterization... The closer the value is to 1, the smaller the attenuation and the closer the error accumulation factor is to 1. When the characterization intensity decreases, the attenuation increases, and the error accumulation factor increases. The initial extrapolated power sequence is gradually compensated for along the extrapolation time using the extrapolation error accumulation factor to obtain the long-term wind power extrapolation result. Gradual compensation refers to correcting the initial extrapolated power value based on the extrapolation error accumulation factor to compensate for the systematic deviation accumulated with the extrapolation step size.

[0058] During wind power extrapolation, as the extrapolation step size increases, the model tends to underestimate the persistence of power or overestimate the decay of power. Therefore, it is necessary to compensate and correct the initial extrapolated power. The compensated long-term wind power extrapolation result is defined as follows: , in, Indicates the extrapolation time The extrapolated results of long-term wind power after compensation This represents the compensation reference benchmark, which is the average power value taken from the last period of historical data. This represents the increment of the error accumulation factor relative to the uncompensated state. The compensation term moderately amplifies the deviation of the extrapolated power from the reference standard as the extrapolation step size increases, offsetting the model's underestimation of long-term power variations. Compensation reference standard. The calculation uses the end of the historical data. The power mean at each time point: , in, This indicates the length of the reference period. The compensation term and the initial extrapolated power value are both in kilowatts, while the error accumulation factor is dimensionless. The extrapolated long-series wind power after compensation remains in kilowatts, ensuring dimensional consistency.

[0059] For all extrapolation times After performing error compensation, the long-term wind power extrapolation results are obtained. .

[0060] For example, error compensation is performed on the initial extrapolated power sequence obtained in step S6. A decay rate coefficient is set. Error cumulative adjustment coefficient For the first extrapolated time... extrapolation step size Calculate the characterization intensity Calculate the extrapolation error accumulation factor Calculate the compensation reference benchmark and set the reference time period length. The average power value of the last 96 moments in the historical data was extracted and calculated. Kilowatts. The initial extrapolated power at the first extrapolation time is known. kilowatts, extrapolated results of long-term wind power calculation after compensation The extrapolated long-series wind power after compensation at the first extrapolation time was 4821.07 kW. For the 48th extrapolation time... extrapolation step size Calculate the characterization intensity Calculate the extrapolation error accumulation factor Assume the initial extrapolation power at the 48th extrapolation time. kilowatts, the result after compensation is: The extrapolated long-series wind power at the 48th extrapolation time was 5256.8 kW. As the extrapolation step size increased, the characterization intensity decreased, the error accumulation factor increased, and the compensation strength correspondingly increased. Error compensation was performed on all 96 extrapolation times to obtain the extrapolated long-series wind power results. .

[0061] S8. After obtaining the long-term wind power extrapolation results, the following is also included: Substitute the extrapolated results of long-term wind power back into the time history of the covariate fusion representation to obtain the extrapolation feedback representation; Calculate the degree of offset between the extrapolated feedback representation and the covariate fusion representation to obtain the representation offset; The extrapolation error accumulation factor is iteratively corrected based on the characterization offset to obtain the corrected extrapolation error accumulation factor, and the long-term wind power extrapolation results are updated using the corrected extrapolation error accumulation factor.

[0062] Specifically, after obtaining the long-term wind power extrapolation results, these results are substituted back into the time history of the covariate fusion representation. The degree of deviation between the extrapolation feedback representation and the covariate fusion representation is calculated. Based on the degree of deviation, the extrapolation error accumulation factor is iteratively corrected, and the corrected extrapolation error accumulation factor is used to update the long-term wind power extrapolation results. This step further improves the extrapolation accuracy through feedback verification between the prediction results and the driving information.

[0063] The extrapolated long-series wind power results are substituted back into the time history of the covariate fusion representation to obtain the extrapolated feedback representation. The extrapolated feedback representation refers to the estimated value of the covariate fusion representation calculated by back-deriving the power results obtained from the extrapolation. There is a driving relationship between the covariate fusion representation and the power; the power extrapolation results carry indirect information about the covariates, and the corresponding representation can be reconstructed through back-mapping.

[0064] Based on the recursive relationship of the power trend components in step S6, a driving relationship is satisfied between power and the covariate fusion representation. The recursive formula is solved in reverse, and the covariate fusion representation is derived from the long-term wind power extrapolation results. The extrapolation feedback representation is obtained by transforming the trend recursive formula: , in, Indicates the extrapolation time Extrapolation feedback representation, and These are the trend autoregressive coefficients and driving coefficients determined in step S6. The feedback characterization is obtained by inverse mapping the compensated power results, carrying the characterization information implicit in the extrapolated power.

[0065] The degree of deviation between the extrapolated feedback representation and the covariate fusion representation is calculated to obtain the representational deviation. The representational deviation refers to the degree of difference between the extrapolated feedback representation and the original covariate fusion representation estimate, reflecting the consistency deviation between the power extrapolation result and the driving information. The representational deviation is small when the extrapolated power highly matches the covariate fusion representation; it is large when the extrapolated power deviates from the driving law of the covariates. The representational deviation over the extrapolation time history is defined as follows: , in, Indicates the extrapolation time The characterization offset, This represents the estimated value of the fused covariate representation obtained through autoregression in step S6. To comprehensively assess the overall shift in the extrapolated time history, the root mean square of the representation shift is calculated: , in, This represents the root mean square of the offset in the extrapolated time history, reflecting the overall offset level. The extrapolation error accumulation factor is iteratively corrected based on the offset to obtain the corrected extrapolation error accumulation factor. Iterative correction means adjusting the error accumulation factor according to the magnitude of the offset. When the offset is large, it indicates that the current error accumulation factor is insufficient to compensate for the extrapolation error, and the error accumulation factor needs to be increased; when the offset is small, it indicates that the compensation is appropriate, and the adjustment range of the error accumulation factor is small. The corrected extrapolation error accumulation factor is defined as follows: , in, Indicates the extrapolation time Corrected extrapolation error accumulation factor, This represents the original extrapolation error accumulation factor in step S7. Indicates the corrected strength coefficient. This represents the absolute value of the offset at the current moment. The root mean square of the offset is represented. This represents a differential constant to prevent the denominator from being zero. It characterizes both the offset and the root mean square as dimensionless quantities; their ratio is also dimensionless. It is a correction intensity coefficient. Since it is a dimensionless quantity, the corrected extrapolation error accumulation factor remains dimensionless, consistent with the dimensions of the original error accumulation factor. The corrected extrapolation error accumulation factor is used to update the long-term wind power extrapolation results. The corrected error accumulation factor is then substituted into the compensation formula in step S7 to recalculate the compensated power extrapolation value: , in, Indicates the extrapolation time Updated long-series wind power extrapolation results, The initial extrapolated power value obtained in step S6, This serves as the compensation reference benchmark in step S7. The feedback correction process can be executed iteratively. In each iteration, the extrapolated feedback representation and representation offset are recalculated using the updated power extrapolation result, and the error accumulation factor is corrected again until the root mean square of the representation offset is reached. Below the convergence threshold Or it may reach the maximum number of iterations. Convergence threshold. The maximum number of iterations is determined based on the normal fluctuation range of the characterization quantity and is used to limit computational overhead.

[0066] For example, feedback correction is performed on the long-term wind power extrapolation results obtained in step S7. Using the 48th extrapolation time... For example, the extrapolated results of the long-term wind power after compensation are known. Kilowatts. Extrapolation feedback representation is calculated using the back-mapping of trend recursion. The trend autoregressive coefficients are known. , driving coefficient kilowatts, and the compensated power values ​​at the first two extrapolated moments. kilowatt, kilowatts. Calculation of extrapolated feedback characterization. The extrapolation feedback representation at the 48th extrapolation time was 0.560.

[0067] The estimated value of the fused covariate representation at this moment, obtained through autoregression in step S6, is known. Calculate the characterization offset. The representation offset at the 48th extrapolated time point is 0.055. The representation offsets are calculated for all 96 extrapolated time points, and the root mean square of the representation offsets is also calculated. It is assumed that the sum of the squares of the representation offsets at each extrapolated time point is... Calculate the root mean square The root mean square of the offset characterizing the extrapolated time history was obtained as 0.044. The extrapolation error accumulation factor at the 48th extrapolation time was corrected based on this offset. A correction strength coefficient was set. Incremental constants Given the original extrapolation error accumulation factor at this moment in step S7. Calculate the corrected extrapolation error accumulation factor. The extrapolation error accumulation factor after correction at the 48th extrapolation time is 1.575.

[0068] The long-term wind power extrapolation results are updated using the corrected error accumulation factor. The initial extrapolated power at the 48th extrapolation time is known. kilowatts, compensation reference benchmark kilowatts. Calculate the updated power extrapolation value. The extrapolated long-series wind power at the 48th extrapolation time was 5330 kW, compared to 5256.8 kW before correction. The compensation strength was enhanced due to the larger characterization offset.

[0069] The extrapolated feedback representation and representation offset are recalculated using the updated power extrapolation results, and the process proceeds to the next iteration. Assume that after the second iteration, the root mean square of the representation offset decreases to 0.021, which is still higher than the set convergence threshold. The third iteration continues; after the third iteration, the root mean square value drops to 0.013, below the convergence threshold, and the iteration terminates. At this point, the converged long-term wind power extrapolation result is obtained. Feedback correction makes the extrapolated power more consistent with the driving law represented by the covariate fusion, improving the overall accuracy of the long-term extrapolation. Feedback correction is performed on all 96 extrapolation times to obtain the final long-term wind power extrapolation result.

[0070] Based on the same inventive concept, such as Figure 2 As shown, the present invention also provides a long-term wind power extrapolation system based on multi-covariate fusion, the system comprising: The sequence acquisition module is used to acquire the time series of multiple meteorological covariates and the historical wind power series of the wind farm to obtain the original multi-covariate sequence; The time-delay alignment module is used to calculate the dynamic time-delay offset of each covariate relative to the historical wind power sequence for each covariate in the original multi-covariate sequence, and to realign each covariate along the time axis based on the dynamic time-delay offset to obtain the time-delay aligned covariate sequence. The coupling metric module is used to measure the coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence within a sliding time window, and to obtain the covariate coupling strength sequence. The covariate fusion module is used to perform weighted fusion of each covariate in the time-delay aligned covariate sequence based on the covariate coupling strength sequence to obtain the covariate fusion representation. The sequence decomposition module is used to decompose the historical wind power sequence into power trend components and power fluctuation components; The stepwise extrapolation module is used to progressively extrapolate and superimpose the power trend component and the power fluctuation component along the extrapolation time history, driven by the covariate fusion representation, to obtain the initial extrapolated power sequence. The error compensation module is used to estimate the extrapolation error accumulation factor based on the degree of attenuation of the covariate fusion representation in the extrapolation time history, and to use the extrapolation error accumulation factor to gradually compensate the initial extrapolation power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

[0071] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A long-term wind power extrapolation method based on multi-covariate fusion, characterized in that, The method includes: The time series of multiple meteorological covariates and the historical wind power series of the wind farm are obtained to obtain the original multi-covariate series; For each covariate in the original multi-covariate sequence, the dynamic time delay offset of each covariate relative to the historical wind power sequence is calculated, and each covariate is realigned along the time axis based on the dynamic time delay offset to obtain the time delay aligned covariate sequence. The coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence is measured within a sliding time window to obtain the covariate coupling strength sequence. Based on the covariate coupling strength sequence, the covariates in the time-delay aligned covariate sequence are weighted and fused to obtain the covariate fusion representation; The historical wind power sequence is decomposed into a power trend component and a power fluctuation component. Driven by covariate fusion representation, the power trend component and the power fluctuation component are progressively extrapolated and superimposed along the extrapolation time history to obtain the initial extrapolated power sequence. The extrapolation error accumulation factor is estimated based on the degree of attenuation of the covariate fusion characterization in the extrapolation time history. The extrapolation error accumulation factor is then used to gradually compensate the initial extrapolated power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

2. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 1, characterized in that, The obtained time-delay aligned covariate sequence includes: Under multiple candidate time delay values, the temporal correlation degree between the shifted subsequence of each covariate and the historical wind power sequence is calculated to obtain the time delay correlation degree distribution; Local extreme values ​​of the time delay correlation distribution are tracked along the time progression direction to obtain the dynamic time delay offset that changes with time. The sampling times of each covariate are shifted and resampled based on the dynamic time delay offset to obtain a time delay aligned covariate sequence.

3. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 2, characterized in that, The obtained covariate coupling strength sequence includes: Within a sliding time window, the fluctuation trends of the time-delay aligned covariate sequence and the historical wind power sequence are extracted to obtain the covariate fluctuation trend and the power fluctuation trend. The consistency of the covariate fluctuation trend and the power fluctuation trend within the same sliding time window is calculated to obtain the in-window consistency measure. The consistency metric within the sliding time window is accumulated and normalized along the advancing direction of the sliding time window to obtain the covariate coupling strength sequence.

4. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 3, characterized in that, The obtained covariate fusion representation includes: Based on the adaptive limit determined by the intensity distribution of the covariate coupling strength sequence, covariates with coupling strength below the adaptive limit in the time-delay aligned covariate sequence are removed to obtain the effective covariate set. Based on the covariate coupling strength sequence, time-varying fusion weights are assigned to each covariate in the effective covariate set, thus obtaining the time-varying fusion weights; The time-varying fusion weights are smoothed and suppressed to obtain stable fusion weights; By superimposing the effective covariate set according to the stable fusion weights at each time step, a covariate fusion representation is obtained.

5. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 4, characterized in that, The process of smoothing and suppressing the time-varying fusion weights to obtain stable fusion weights includes: Applying a smoothing constraint to the time-varying fusion weights along the time direction yields smoothed fusion weights; Calculate the magnitude of change in the smoothed fusion weights between adjacent time points to obtain the weight abrupt change degree; When the weight abruptness exceeds the allowable range determined by the distribution of weight abruptness changes, the smoothed fusion weights are pulled back to suppress them, resulting in stable fusion weights.

6. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 4, characterized in that, The process of decomposing the historical wind power sequence into power trend components and power fluctuation components includes: Multi-timescale smoothing was performed on the historical wind power series to obtain smoothed series at different time scales; The power trend component is obtained by extracting the gradient component from smooth sequences at different time scales. The power fluctuation component is obtained by subtracting the historical wind power sequence from the power trend component over time.

7. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 6, characterized in that, The process of obtaining the initial extrapolated power sequence includes: Driven by covariate fusion characterization, the power trend component is progressively extrapolated along the extrapolation time history to obtain the trend extrapolation component. Driven by covariate fusion characterization, the power fluctuation component is progressively extrapolated along the extrapolated time history to obtain the fluctuation extrapolation component. The trend extrapolation component and the fluctuation extrapolation component are aligned and superimposed according to the extrapolation time to obtain the initial extrapolation power sequence.

8. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 7, characterized in that, The obtained long-time series wind power extrapolation results include: Based on the change in the intensity of the characterization along the extrapolated time history using the covariate fusion characterization, the characterization decay curve is obtained; Based on the degree to which the decay curve decreases with the extrapolation step size, the extrapolation error accumulation factor is obtained; The extrapolation error accumulation factor is used to compensate the initial extrapolated power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.

9. The long-term wind power extrapolation method based on multi-covariate fusion according to claim 8, characterized in that, After obtaining the long-term wind power extrapolation results, the following are also included: Substitute the extrapolated results of long-term wind power back into the time history of the covariate fusion representation to obtain the extrapolation feedback representation; Calculate the degree of offset between the extrapolated feedback representation and the covariate fusion representation to obtain the representation offset; The extrapolation error accumulation factor is iteratively corrected based on the characterization offset to obtain the corrected extrapolation error accumulation factor, and the long-term wind power extrapolation results are updated using the corrected extrapolation error accumulation factor.

10. A long-term wind power extrapolation system based on multi-covariate fusion, characterized in that, The system includes: The sequence acquisition module is used to acquire the time series of multiple meteorological covariates and the historical wind power series of the wind farm to obtain the original multi-covariate sequence; The time-delay alignment module is used to calculate the dynamic time-delay offset of each covariate relative to the historical wind power sequence for each covariate in the original multi-covariate sequence, and to realign each covariate along the time axis based on the dynamic time-delay offset to obtain the time-delay aligned covariate sequence. The coupling metric module is used to measure the coupling strength between each covariate in the time-delay aligned covariate sequence and the historical wind power sequence within a sliding time window, and to obtain the covariate coupling strength sequence. The covariate fusion module is used to perform weighted fusion of each covariate in the time-delay aligned covariate sequence based on the covariate coupling strength sequence to obtain the covariate fusion representation. The sequence decomposition module is used to decompose the historical wind power sequence into power trend components and power fluctuation components; The stepwise extrapolation module is used to progressively extrapolate and superimpose the power trend component and the power fluctuation component along the extrapolation time history, driven by the covariate fusion representation, to obtain the initial extrapolated power sequence. The error compensation module is used to estimate the extrapolation error accumulation factor based on the degree of attenuation of the covariate fusion representation in the extrapolation time history, and to use the extrapolation error accumulation factor to gradually compensate the initial extrapolation power sequence along the extrapolation time to obtain the long-term wind power extrapolation result.