A method and system for predicting ultra-short-term wind power using a dual-path parallel fusion approach

By employing a dual-path parallel fusion method for wind power ultra-short-term power prediction, utilizing a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path, combined with multi-scale input data and bias correction, the method solves the prediction bias and response lag problems of existing wind power ultra-short-term power prediction under complex wind conditions, achieving higher prediction accuracy and stability.

CN123121497APending Publication Date: 2026-11-17BEIJING EAST ENVIRONMENT ENERGY TECH +6
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611270169.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing wind power ultra-short-term power prediction methods have low prediction accuracy and slow response in scenarios such as drastic wind speed fluctuations, power ramp-ups, and oscillation transitions. They are also unable to characterize the local morphology and periodic features of the power sequence, leading to amplified prediction errors.

Method used

A dual-path parallel fusion wind power ultra-short-term power prediction method is adopted. By constructing a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path, data-driven and physical rule-based prediction is performed using multi-scale input data. Combined with time-series prediction values ​​and deviation correction reference values, ensemble prediction is performed to improve the stability and accuracy of the prediction results.

Benefits of technology

Under complex wind conditions, it improves the accuracy and stability of ultra-short-term wind power forecasting, reduces forecast bias, and enhances the ability to characterize local morphology and periodic features.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121497A_ABST
    Figure CN123121497A_ABST
Patent Text Reader

Abstract

This application provides a dual-path parallel fusion method and system for wind power ultra-short-term power prediction. The method acquires wind farm operation data and multi-source numerical weather prediction data to construct multi-scale input data. It constructs two parallel prediction paths: a locally attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path. The locally attention-enhanced lightweight linear prediction path is trained using sequence morphology preservation loss based on time-domain local statistics and frequency-domain amplitude spectrum, and outputs data-driven prediction results. The hierarchical adaptive prediction path determines the prediction strategy based on historical power sequences, fuses the time-series prediction values ​​with a deviation correction reference value determined based on historical prediction deviation statistics, and outputs a physical rule prediction result. The two prediction results are then combined for ensemble prediction to obtain the wind power ultra-short-term power prediction sequence. This method considers both physical evolution and sequence morphology, and can reduce prediction errors in scenarios with sudden wind speed changes and power ramp-up.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and more specifically, to a method and system for predicting wind power ultra-short-term power using a dual-path parallel fusion approach. Background Technology

[0002] Wind farm output power is affected by wind speed variations. Drastic wind speed fluctuations, power ramp-ups, and oscillation transitions can cause the power sequence to exhibit non-stationary and rapidly changing characteristics. Ultra-short-term wind power forecasting needs to reflect power change trends and sequence dynamics to establish an ultra-short-term power benchmark for wind farms.

[0003] Existing solutions mostly employ a single data-driven model for prediction, using a point-to-point loss function to independently constrain the prediction results at each time point. In scenarios involving drastic wind speed fluctuations, power ramp-ups, and oscillation transitions, the single data-driven model suffers from low prediction accuracy and response lag, easily amplifying prediction errors due to trend misjudgment. Furthermore, the point-to-point loss function focuses on point-by-point errors, failing to characterize the local morphology and periodic features of the power sequence, leading to morphological deviations between the predicted and measured waveforms. In summary, existing solutions suffer from insufficient predictive adaptability and inadequate sequence morphological constraints under complex wind conditions.

[0004] Therefore, this application provides a dual-path parallel fusion wind power ultra-short-term power prediction method and system to solve one of the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for wind power ultra-short-term power prediction using a dual-path parallel fusion approach, which can solve at least one of the aforementioned technical problems. The specific solution is as follows: According to a specific embodiment of this application, in a first aspect, this application provides a dual-path parallel fusion method for wind power ultra-short-term power prediction, comprising: Wind farm operation data and multi-source numerical weather prediction data are acquired to construct multi-scale input data. A local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path are constructed for parallel processing. The local attention-enhanced lightweight linear prediction path is trained based on predicted power sequences and actual power sequences, which are constrained by time-domain local statistics and frequency-domain amplitude spectrum sequence shape preservation loss. The multi-scale input data is input into the local attention-enhanced lightweight linear prediction path to obtain data-driven prediction results. Based on historical power sequences in the multi-scale input data, a prediction strategy is determined, and time-series predicted values ​​and deviation correction reference values ​​are fused based on the hierarchical adaptive prediction path to obtain physical rule prediction results. The time-series predicted values ​​are determined based on the prediction strategy, and the deviation correction reference values ​​are determined based on historical prediction deviation statistics. The data-driven prediction results and the physical rule prediction results are then combined for ensemble prediction to obtain a wind power ultra-short-term power prediction sequence.

[0006] According to a specific embodiment of this application, in a second aspect, this application provides a dual-path parallel fusion wind power ultra-short-term power prediction system, comprising: A data construction module is used to acquire wind farm operation data and multi-source numerical weather prediction data to construct multi-scale input data; it constructs a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path for parallel processing; wherein the local attention-enhanced lightweight linear prediction path is trained based on the predicted power sequence and the actual power sequence, and the predicted power sequence and the actual power sequence are constrained by the sequence shape preservation loss of the time domain local statistics and the frequency domain amplitude spectrum; a dual-path prediction module is used to input the multi-scale input data into the local attention-enhanced lightweight linear prediction path to obtain data-driven prediction results; based on the historical power sequence in the multi-scale input data, a prediction strategy is determined, and the time-series prediction value and the deviation correction reference value are fused to obtain the physical rule prediction result; wherein the time-series prediction value is determined based on the prediction strategy, and the deviation correction reference value is determined based on the historical prediction deviation statistics; a prediction output module is used to perform ensemble prediction on the data-driven prediction result and the physical rule prediction result to obtain the wind power ultra-short-term power prediction sequence.

[0007] Compared with the prior art, the above-described solutions of this application have at least the following beneficial effects: This application provides a dual-path parallel fusion method for ultra-short-term wind power prediction. This method constructs multi-scale input data from wind farm operation data and multi-source numerical weather prediction data, fusing operational and meteorological information. A local attention-enhanced lightweight linear prediction path uses time-domain local statistics and frequency-domain amplitude spectra to constrain the predicted and actual power sequences, ensuring that data-driven prediction results take into account both local morphology and periodic characteristics. A hierarchical adaptive prediction path determines the prediction strategy based on historical power sequences and fuses time-series predicted values ​​with bias correction reference values, enabling physical rule-based prediction results to adapt to different wind conditions and suppress bias amplification. An ensemble prediction is performed on the data-driven and physical rule-based prediction results, leveraging the complementarity of the two predictions to improve the stability and accuracy of prediction results under scenarios of severe wind speed fluctuations and power ramp-up. Attached Figure Description

[0008] Figure 1 A flowchart illustrating a dual-path parallel fusion wind power ultra-short-term power prediction method provided in some embodiments of this application; Figure 2 A flowchart illustrating sequence morphology preservation loss determination and local attention-enhanced lightweight linear prediction path training provided in some embodiments of this application; Figure 3 A general flowchart for constructing multi-scale input data provided in some embodiments of this application; Figure 4 Flowcharts illustrating the construction of derived features and the generation of candidate feature sets provided in some embodiments of this application; Figure 5 Flowcharts for generating basic feature data provided in some embodiments of this application; Figure 6 Flowcharts for filtering key input feature subsets provided in some embodiments of this application; Figure 7 Flowcharts illustrating the multi-scale organization of key input feature subsets provided in some embodiments of this application; Figure 8 Flowcharts of local attention-enhanced lightweight linear prediction paths provided in some embodiments of this application; Figure 9 Flowcharts for determining prediction strategies based on cascaded short and long windows provided in some embodiments of this application; Figure 10 A flowchart illustrating the determination of time-series predicted values ​​based on a prediction strategy, provided for some embodiments of this application; Figure 11 A flowchart illustrating the fusion of time-series predicted values ​​and bias correction reference values ​​provided in some embodiments of this application; Figure 12The diagram shows the structure of a dual-path parallel fusion wind power ultra-short-term power prediction system provided in some embodiments of this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0011] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0012] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0013] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0014] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a product or system comprising a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a product or system. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the product or system that includes said element.

[0015] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0016] The optional embodiments of this application are described in detail below with reference to the accompanying drawings.

[0017] The embodiments provided in this application are embodiments of a dual-path parallel fusion wind power ultra-short-term power prediction method.

[0018] The following is combined Figure 1 The embodiments of this application will be described in detail.

[0019] Figure 1 A flowchart of a dual-path parallel fusion wind power ultra-short-term power prediction method provided in some embodiments of this application.

[0020] Reference Figure 1 The dual-path parallel fusion wind power ultra-short-term power prediction method includes the following steps: S101, acquire wind farm operation data and multi-source numerical weather forecast data to construct multi-scale input data; S102, Construct a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path for parallel processing; Among them, the local attention-enhanced lightweight linear prediction path is trained based on the predicted power sequence and the real power sequence, and the predicted power sequence and the real power sequence are constrained by the sequence shape preservation loss based on the local statistics in the time domain and the amplitude spectrum in the frequency domain. S103a: Input multi-scale input data into a local attention-enhanced lightweight linear prediction path to obtain data-driven prediction results; S103b: Based on the historical power sequence in the multi-scale input data, a prediction strategy is determined, and the time-series prediction value is fused with the bias correction reference value based on the hierarchical adaptive prediction path to obtain the physical rule prediction result. Among them, the time-series forecast value is determined based on the forecasting strategy, and the deviation correction reference value is determined based on the historical forecast deviation statistics. S104, perform ensemble prediction on the data-driven prediction results and the physical rule prediction results to obtain the wind power ultra-short-term power prediction sequence.

[0021] In this embodiment, the local attention-enhanced lightweight linear prediction path extracts temporal features from multi-scale input data and forms data-driven prediction results. The hierarchical adaptive prediction path uses historical power sequences from multi-scale input data to determine wind conditions and perform temporal extrapolation, and combines them with deviation correction reference values ​​to form physical rule prediction results.

[0022] For example, the locally attention-enhanced lightweight linear prediction path is trained on both the predicted power sequence and the true power sequence before performing predictions. The sequence morphology preservation loss uses temporal local statistics to constrain correlation, relative scale, and mean level on the corresponding local segments, and uses frequency domain amplitude spectrum to constrain the periodic components on the complete sequence.

[0023] For example, the trained local attention-enhanced lightweight linear prediction path takes multi-scale input data and outputs data-driven prediction results.

[0024] For example, the hierarchical adaptive prediction path obtains historical power sequences from multi-scale input data, determines the prediction strategy based on the local fluctuations and reversal relationships of the historical power sequences, and finally forms time-series prediction values ​​based on the determined prediction strategy.

[0025] In some embodiments, the deviation correction reference value is determined based on the statistical results of historical prediction deviations, and the hierarchical adaptive prediction path fuses the time-series prediction value with the deviation correction reference value to obtain the physical rule prediction result.

[0026] For example, ensemble prediction can be performed using parallel ensemble or hierarchical fusion.

[0027] As a feasible implementation method, a parallel integration approach is adopted to aggregate predictions by voting or averaging the independent prediction results of two prediction paths.

[0028] As another feasible implementation method, a hierarchical fusion approach is adopted to aggregate predictions. The meta-learner learns the output distribution characteristics of the two prediction results and performs secondary calibration on the two prediction results to output the wind power ultra-short-term power prediction sequence.

[0029] Figure 2 This is a flowchart illustrating the sequence morphology preservation loss determination and local attention enhancement lightweight linear prediction path training provided in some embodiments of this application.

[0030] Reference Figure 2 This includes the following steps: S201, obtain the historical window multi-scale input data and the future window actual power sequence of the same training sample; S202, input the historical window multi-scale input data into the local attention-enhanced lightweight linear prediction path to obtain the predicted power sequence, and determine the actual power sequence of the future window as the true power sequence; S203, a sliding window is used to divide the predicted power sequence and the true power sequence into multiple predicted local segments and multiple true local segments respectively, and a segment correspondence is established according to the position of the predicted local segments and the true local segments in the corresponding sequences; S204. Based on the Pearson correlation coefficient between the predicted local segments and the true local segments with segment correspondence, determine the local correlation loss; remove the means of the predicted local segments and the true local segments respectively, map them to the predicted probability distribution and the true probability distribution through a normalized exponential function, and determine the relative scale loss based on the Jensen-Shannon divergence between the predicted probability distribution and the true probability distribution with segment correspondence; determine the mean shift loss based on the absolute error between the means of the predicted local segments and the true local segments with segment correspondence; perform discrete Fourier transforms on the predicted power sequence and the true power sequence respectively, and determine the frequency domain amplitude matching loss based on the mean square error between the amplitude spectra of the two transform results. S205, the local correlation loss, relative scale loss, mean shift loss and frequency domain amplitude matching loss are weighted and summed to obtain the sequence shape preservation loss, and the parameters of the local attention-enhanced lightweight linear prediction path are updated according to the sequence shape preservation loss.

[0031] In this embodiment of the application, each training sample includes multi-scale input data of a fixed-length historical window, and the actual power sequence of the future window corresponding to the historical window.

[0032] In one specific implementation, the historical window and the future window are configured to have the same length, so that each predicted position in the predicted power sequence corresponds to each actual power position in the true power sequence.

[0033] In some embodiments, a sliding window can be used to move across the predicted power sequence and the actual power sequence respectively, dividing the two sequences into multiple local segments, and establishing a segment correspondence relationship based on the position of each local segment in the corresponding sequence.

[0034] In one specific implementation, the step size of the sliding window can be half the length of the sliding window.

[0035] For example, the sequence morphology preservation loss is formed by a weighted average of local correlation loss, relative scale loss, mean shift loss, and frequency domain amplitude matching loss, calculated as follows: ; in, This indicates the loss in sequence shape preservation. This indicates the loss of local correlation. Indicates relative scale loss. This represents the loss due to mean shift. This represents the frequency domain amplitude matching loss. , , and This represents the structure regularization coefficient.

[0036] For example, for each group of predicted local segments and real local segments with segment correspondence, the Pearson correlation coefficient between them is calculated. The result of subtracting the Pearson correlation coefficient is taken as the correlation loss of the corresponding segment. Then, the correlation loss of all corresponding segments is averaged to obtain the local correlation loss.

[0037] Specifically, the calculation method is expressed as follows: ; in, This indicates the loss of local correlation. Indicates the number of local segments, Indicates the index of a local segment. Indicates the first A real local segment, Indicates the first A predicted local segment, Indicates the first The first real local segment and the first Pearson correlation coefficient between predicted local segments.

[0038] In the above embodiments, the local correlation loss constrains the predicted local segments and the real local segments with corresponding relationships in terms of rising and falling trends and peak order.

[0039] For example, for each corresponding local segment, the mean of the predicted local segment and the mean of the true local segment are subtracted respectively to eliminate the influence of the overall level on the local relative scale comparison. Then, the two mean-reduced local segments are mapped to the predicted probability distribution and the true probability distribution using a normalized exponential function. The Jensen-Shannon divergence between the two probability distributions is calculated, and the divergence of all corresponding segments is averaged to obtain the relative scale loss. ; in, Indicates relative scale loss. Indicates the number of local segments, Indicates the index of a local segment. Indicates the Jansen-Shannon divergence. Represents the normalized exponential function, Indicates the first A real local segment, Indicates the first The mean of a real local segment, Indicates the first A predicted local segment, Indicates the first The mean of each predicted local segment, This represents the separation relationship between two probability distributions in the Jensen-Shannon divergence.

[0040] In this embodiment, the relative scale loss is used to constrain the relative relationship of fluctuation amplitudes in the corresponding local segments, reducing the possibility of over-smoothing or oscillation mismatch in the predicted power sequence.

[0041] For example, for each corresponding local segment, calculate the absolute error between the mean of the actual local segment and the mean of the predicted local segment, and then average the absolute errors of all corresponding segments to obtain the mean shift loss: ; in, This represents the loss due to mean shift. Indicates the number of local segments, Indicates the index of a local segment. Indicates the first The mean of a real local segment, Indicates the first The mean of each predicted local segment.

[0042] In this embodiment, the mean offset loss is used to correct the overall horizontal drift of the predicted local segment relative to the true local segment.

[0043] For example, discrete Fourier transforms are performed on the complete predicted power sequence and the true power sequence respectively, the amplitude spectra of the two transform results are obtained, and the mean square error between the two amplitude spectra is calculated to obtain the frequency domain amplitude matching loss: ; in, This represents the frequency domain amplitude matching loss. Indicates the lengths of the predicted power sequence and the actual power sequence. Represents the Discrete Fourier Transform. Represents the true power sequence. Represents the predicted power sequence. In Denotes the square norm 2. In This indicates taking the modulo of each element.

[0044] In this embodiment, the frequency domain amplitude matching loss is used to constrain the reproduction of the dominant periodic components and energy distribution in the real power sequence by the predicted power sequence, and to suppress spurious frequency components.

[0045] In the above embodiments, the three temporal loss components of sequence morphology preservation loss can be determined in parallel after the segment correspondence is established, and the frequency domain amplitude matching loss can also be determined in parallel with the three temporal loss components.

[0046] For example, after obtaining the four losses, they can be weighted and summed according to the corresponding structure regularization coefficients, and the parameters of the local attention-enhanced lightweight linear prediction path can be updated according to the morphology-preserving loss of the obtained sequence.

[0047] In some embodiments, a local attention-enhanced lightweight linear prediction path can be trained in an end-to-end manner.

[0048] For example, in each training round, training samples are loaded in mini-batch mode, and forward propagation, loss calculation, gradient backpropagation, and parameter update are performed sequentially.

[0049] In some embodiments, an adaptive moment estimation optimizer with weight decay can be used for parameter updates, and a cosine annealing scheduling strategy can be used to dynamically adjust the learning rate during training.

[0050] For example, after each training round, the loss value for the current path can be calculated on the validation data.

[0051] Furthermore, if the verification loss does not decrease for a preset number of consecutive rounds, training is terminated, and the path parameters corresponding to the lowest verification loss are restored.

[0052] In addition, the validation data can be used to perform a grid search on the structure regularization coefficients of the four losses and select coefficient combinations according to the principle of minimizing validation loss.

[0053] For example, after training and validation are complete, the saved path parameters are used to evaluate on the test data.

[0054] In this embodiment, by feeding multi-scale input data to parallel data-driven prediction paths and physical rule prediction paths, and maintaining the predicted power sequence of the loss-constrained data-driven path in sequence form, and then performing ensemble prediction on the prediction results generated by the two paths, it is possible to simultaneously utilize the temporal features in multi-source data, the wind condition change relationship corresponding to historical power sequences, and the local morphology and periodic components of the prediction sequence, thereby reducing the possibility of misjudgment of single-path trends and amplification of prediction deviations, and improving the stability and accuracy of wind power ultra-short-term power prediction sequences under scenarios of severe wind speed fluctuations and power ramp-up.

[0055] Figure 3 A general flowchart for constructing multi-scale input data provided in some embodiments of this application.

[0056] Reference Figure 3 This includes the following steps: S301, acquire wind farm operation data and multi-source numerical weather forecast data; S302 performs time dimension alignment, outlier sample processing, and normalization on wind farm operation data and multi-source numerical weather forecast data to obtain basic feature data; S303, construct derived features based on basic feature data, combine basic feature data and derived features to obtain candidate feature set, and select key input feature subset from candidate feature set; S304 organizes the key input feature subsets into multi-scale data to obtain multi-scale input data.

[0057] In this embodiment, the key input feature subset is organized by time scale, spatial scale and meteorological source scale to obtain multi-scale input data for use by two prediction paths.

[0058] Figure 4 This is a flowchart illustrating the construction of derived features and the generation of candidate feature sets provided in some embodiments of this application.

[0059] Reference Figure 4 This includes the following steps: S401, Extract processing objects from basic feature data to construct various derived features; S402: Turbulence intensity characteristics are formed based on the wind speeds at each vertical layer of the meteorological tower and the nacelle wind speed; observation bias characteristics and wind speed difference characteristics between meteorological sources are formed based on the numerical weather forecast wind speeds at the matching height of each meteorological source, the actual wind speeds measured at the meteorological tower, and the numerical weather forecast wind speeds of different meteorological sources; the wind direction angles at each vertical layer of the meteorological tower and each vertical layer of each meteorological source are converted into Cartesian components to form wind direction decomposition characteristics; lag characteristics are formed based on the actual grid-connected power, nacelle wind speed, and the numerical weather forecast wind speeds of each meteorological source at the corresponding time; and sliding statistical characteristics are formed based on the nacelle wind speed, actual grid-connected power, and key vertical layer data of each meteorological source within the sliding window. S403 uses turbulence intensity characteristics, observation bias characteristics, wind speed difference characteristics between meteorological sources, wind direction decomposition characteristics, hysteresis characteristics, and sliding statistics characteristics as various derived characteristics. The basic feature data is combined with various derived characteristics to obtain a candidate feature set.

[0060] For example, wind speeds and nacelle wind speeds are extracted from the basic feature data for each vertical layer of the wind measurement tower. For any extracted wind speed data, a sliding window consisting of a preset number of time steps is determined. The standard deviation and mean of the wind speed within the sliding window are calculated, and the turbulence intensity characteristics are determined based on the ratio of the sum of the standard deviation, the mean, and the division-to-zero constant. ; in, This indicates the turbulence intensity at the corresponding time and the source of the corresponding wind speed. This indicates the wind speed source corresponding to the wind speed at a vertical level of the meteorological tower or in the nacelle. This represents the standard deviation of wind speed within the sliding window. This represents the average wind speed within the sliding window. This represents the smallest positive number that is protected against division by zero. Indicates time, Indicates wind speed. This indicates a sliding window.

[0061] In this embodiment, turbulence intensity features are used to quantify the degree of fluctuation of the corresponding wind speed data within the sliding window and to provide input information related to instantaneous power fluctuations for subsequent models.

[0062] For example, the numerical weather forecast wind speed and the wind speed measured by the anemometer tower at the matching height of each meteorological source are extracted from the basic feature data. The difference between the two wind speeds is used to obtain the observation bias characteristics:

[0063] in, This indicates that the vertical coordinate layer of the numerical weather prediction system matches the vertical coordinate layer of the wind measurement tower. Indicates time Lower meteorological source In the vertical coordinate layer Numerical weather forecast wind speed and vertical layers of the wind measurement tower The observed deviation between the measured wind speed and the actual wind speed Indicates time Lower meteorological source In the vertical coordinate layer Numerical weather forecast wind speed, Indicates time Lower wind tower vertical layer The measured wind speed, Indicates time, Indicates meteorological source index, This represents the vertical coordinate layer corresponding to the numerical weather forecast variables. This indicates the vertical layer corresponding to the wind measurement tower sensor.

[0064] Furthermore, the wind speed difference between different meteorological sources is obtained by subtracting the numerical weather forecast wind speeds at corresponding times and corresponding vertical layers.

[0065] In this embodiment of the application, the observation bias feature characterizes the deviation between the numerical weather forecast wind speed and the measured wind speed, and the wind speed difference feature between meteorological sources characterizes the forecast dispersion between different meteorological sources.

[0066] For example, wind direction angles are extracted from each vertical layer of the wind measurement tower and each vertical layer of each meteorological source from the basic feature data, and the wind direction angles are converted into two Cartesian components: The first Cartesian component is calculated as follows: ; in, This represents the first Cartesian component corresponding to the wind direction angle. This indicates the wind direction angle in degrees.

[0067] The second Cartesian component is calculated as follows: ; in, This represents the second Cartesian component corresponding to the wind direction angle. This indicates the wind direction angle in degrees.

[0068] In this embodiment, two Cartesian components together constitute the wind direction decomposition feature, and are used for wind direction data of each vertical layer of the wind measurement tower and each vertical layer of the meteorological source.

[0069] In some embodiments, when constructing lag features, the actual grid-connected power and cabin wind speed of a preset time step prior to the current moment can be extracted from the basic feature data, and the numerical weather forecast wind speed of each meteorological source at the corresponding moment can be added to obtain lag features, which are used to characterize the temporal relationship between preceding power, preceding cabin wind speed and corresponding forecast wind speed.

[0070] The lag time step is a specific implementation parameter, which is set according to the historical window used.

[0071] In some embodiments, when constructing sliding statistical features, the sliding average and sliding standard deviation of the cabin wind speed, the sliding average of the actual grid-connected power, and the sliding average and sliding standard deviation of the key vertical layer data of each meteorological source can be calculated separately within a sliding window formed by a preset number of time steps in the past. The resulting sliding statistics together form sliding statistical features.

[0072] For example, wind farm operation data includes actual grid-connected power, available power, nacelle wind speed, and wind speed, wind direction, temperature, and relative humidity at each vertical level of the meteorological tower.

[0073] For example, multi-source numerical weather forecast data includes numerical weather forecasts of wind speed, wind direction, temperature, humidity, and precipitation from different meteorological sources at different vertical layers.

[0074] In some embodiments, each vertical layer of the wind measurement tower corresponds to a sensor installed at a different altitude, and different vertical layers in the numerical weather prediction data correspond to different vertical coordinate layers of the numerical weather prediction output variables.

[0075] In this embodiment of the application, different meteorological sources correspond to forecast products from different meteorological centers.

[0076] In one specific implementation, the vertical layers of the wind measurement tower may include layers at heights of 10 meters, 30 meters, 50 meters, and 70 meters.

[0077] In some specific embodiments, basic feature data can be obtained by preprocessing and cleaning wind farm operation data and multi-source numerical weather forecast data.

[0078] Figure 5 A flowchart illustrating the generation of basic feature data provided for some embodiments of this application.

[0079] Reference Figure 5 This includes the following steps: S501 performs inner connection and alignment on wind farm operation data and multi-source numerical weather forecast data to obtain equally spaced time series samples. S502 divides the equally spaced time series samples into training data, validation data, and test data; S503a uses a statistical threshold based on the interquartile range to judge the anomalies of each continuous feature in the training data and obtain the first candidate abnormal sample. S503b standardizes each continuous feature in the training data to obtain the anomaly detection feature space; S504a, K-means clustering is used in the anomaly detection feature space to determine anomalies and obtain the second candidate anomaly sample; S504b uses isolated forest to judge anomalies in the anomaly detection feature space and obtains the third candidate anomaly sample. S505, remove the intersection of the first candidate abnormal sample, the second candidate abnormal sample and the third candidate abnormal sample from the training data to obtain the cleaned training data; S506. Based on the cleaned training data, determine the minimum and maximum values ​​of each feature, and perform normalization processing on the cleaned training data, validation data, and test data based on the minimum and maximum values ​​to obtain the basic feature data.

[0080] As one specific implementation, the time resolution of the time series samples can be set to 15 minutes, and the resulting data can be used to predict the power in the next 15 minutes to 4 hours.

[0081] In one specific implementation, training data, validation data, and test data account for 60%, 20%, and 20% of the complete time series, respectively. Among them, training data is used for path parameter learning, validation data is used for hyperparameter tuning and early stopping monitoring, and test data is used for evaluation after training and validation are completed.

[0082] For example, in the anomaly detection branch based on statistical thresholds, for each continuous feature in the training data, the first and third quartiles can be determined using the following formula, and the interquartile range can be calculated: ; in, Indicates the interquartile range. Indicates the first quartile. This represents the third quartile.

[0083] As one specific implementation method, the normal numerical range determined based on the interquartile range can be represented as follows: ; in, Indicates the first quartile. Indicates the third quartile. This represents the interquartile range.

[0084] In one specific implementation, training samples whose values ​​exceed the normal range are marked as first candidate abnormal samples.

[0085] For example, in another anomaly detection branch, the continuous features in the training data are standardized to obtain an anomaly detection feature space. Using this feature space as a common input, K-means clustering and isolated forest are used for anomaly detection, respectively, to obtain second and third candidate anomaly samples. K-means clustering and isolated forest are used concurrently, without one anomaly detection result being used as input for the other.

[0086] For example, after obtaining the first candidate abnormal sample, the second candidate abnormal sample, and the third candidate abnormal sample, the intersection of the three candidate abnormal samples is determined, and only the samples that are judged as abnormal by all three abnormality judgment methods are removed from the training data to obtain the cleaned training data.

[0087] For example, the minimum and maximum values ​​of each feature are determined based on the cleaned training data, and the obtained minimum and maximum values ​​are saved. For any feature, the calculation relationship for min-max normalization is as follows: ; in, Representation of features Normalized eigenvalues Representation of features Eigenvalues ​​before normalization This represents the features determined based on the cleaned training data. The minimum value, This represents the features determined based on the cleaned training data. The maximum value, Indicates feature index, This represents the cleaned training data used to determine the normalization parameters.

[0088] In this embodiment, the same set of minimum and maximum values ​​determined from the cleaned training data are used to normalize the cleaned training data, validation data, and test data respectively, to obtain basic feature data. Specifically, the intersection of the three anomaly detection results is taken before removing samples, which reduces sample removal bias caused by a single anomaly detection. Furthermore, determining the normalization parameters only based on the cleaned training data and using the same parameters for the validation and test data avoids information from the validation and test data from being included in the normalization parameters.

[0089] Figure 6 A flowchart illustrating the filtering of key input feature subsets provided for some embodiments of this application.

[0090] Reference Figure 6 This includes the following steps: S601a, Input the candidate feature set into the random forest model to obtain the random forest prediction results; S601b: Input the candidate feature set into the extreme gradient boosting model to obtain the extreme gradient boosting prediction results; S602a, based on the average reduction in Gini impurity caused by the splitting of each decision tree node in the random forest model, the first feature importance of each candidate feature is obtained; S602b: Based on the average gain contribution of each split point in the extreme gradient boosting model, the second feature importance of each candidate feature is obtained; S602c, performs Shapley additive interpretation on the random forest prediction results or extreme gradient boosting prediction results to obtain the importance of the third feature of each candidate feature; S603, normalize the importance of the first feature, the second feature, and the third feature of each candidate feature and sum them by weight to obtain the comprehensive score of each candidate feature; S604: Sort the candidate features from high to low according to the comprehensive score, and select the candidate features in turn until the cumulative score of the selected candidate features exceeds the preset threshold, thus obtaining a subset of key input features.

[0091] In some specific embodiments, for each candidate feature in the candidate feature set, the amount of Gini impurity reduction generated by the candidate feature when used for node splitting in each decision tree of the random forest model is counted, and the first feature importance of the candidate feature is obtained based on the average result of the corresponding Gini impurity reduction.

[0092] In some specific embodiments, the gain contribution of the candidate feature when it is used for splitting in the Extreme Gradient Boosting (XGBoost) model is statistically analyzed, and the second feature importance of the candidate feature is obtained based on the average result of the corresponding gain contribution.

[0093] In some specific embodiments, the random forest prediction results or extreme gradient boosting prediction results are subjected to Shapley additive interpretation (SHAP) to determine the contribution of the candidate feature to the model prediction results, and the third feature importance of the candidate feature is obtained according to the corresponding contribution.

[0094] In this embodiment, the importance of the first feature, the importance of the second feature, and the importance of the third feature of each candidate feature are normalized respectively, and then the three normalized feature importances are weighted and summed to obtain the comprehensive score of the corresponding candidate feature. The specific calculation method is as follows: ; in, Indicates the first The combined score of each candidate feature Indicates the first The importance of the first feature after normalization of the candidate features Indicates the first The importance of the second feature after normalization of the candidate features Indicates the first The importance of the third feature after normalization of the candidate features , and This represents the weights corresponding to the importance of the three normalized features. Indicates the candidate feature index. This represents the random forest model. This represents an extreme gradient boosting model.

[0095] In this embodiment, candidate features are sorted from highest to lowest based on their comprehensive scores, and candidate features are selected sequentially according to the sorting results, with the corresponding comprehensive scores accumulated. Further, if the accumulated score does not exceed a preset threshold, the next candidate feature is selected; if the accumulated score exceeds the preset threshold, the selection stops, and the selected candidate features are combined into a subset of key input features.

[0096] The embodiments of this application combine a random forest model, an extreme gradient boosting model, and a Shapley additive interpretation to form a comprehensive score, which can reduce the feature importance bias formed by a single model and improve the robustness and interpretability of the key input feature selection results.

[0097] Figure 7 A flowchart illustrating the multi-scale organization of key input feature subsets provided for some embodiments of this application.

[0098] Reference Figure 7 This includes the following steps: S701a: Extract current time features, lag features, and sliding statistical features with the same timestamp from the key input feature subset, and superimpose the current time features, lag features, and sliding statistical features to obtain time scale features; S701b extracts wind tower layered data, cabin height data, and numerical weather prediction multi-vertical layer data from key input feature subsets to determine vertical inter-layer difference features, and integrates wind tower layered data, cabin height data, numerical weather prediction multi-vertical layer data, and vertical inter-layer difference features to obtain spatial scale features; S701c organizes the forecast data, wind speed difference characteristics between meteorological sources, and observation bias characteristics of each meteorological source in the key input feature subset into meteorological source scale features. S702, standardize the time scale characteristics, spatial scale characteristics and meteorological source scale characteristics by standard scores to obtain standardized time scale characteristics, standardized spatial scale characteristics and standardized meteorological source scale characteristics; S703 splices standardized time-scale features, standardized spatial-scale features, and standardized meteorological source-scale features along the feature dimensions to obtain multi-scale input data.

[0099] For example, when forming time-scale features, the current moment features, lag features, and sliding statistical features are extracted from the key input feature subset according to the timestamp, and the three types of features with the same timestamp are superimposed.

[0100] As one specific implementation method, the original data adopts a 15-minute time resolution, and the lag characteristics and sliding statistical characteristics of the preceding time are superimposed on this time resolution, so that the time scale characteristics simultaneously include the current time information, the time lag information, and the window statistical information.

[0101] For example, when forming spatial scale features, wind tower layered data, cabin height data, and numerical weather prediction multi-vertical layer data are extracted from the key input feature subset. Based on the extracted data, the vertical inter-layer difference features are determined, and then the wind tower layered data, cabin height data, numerical weather prediction multi-vertical layer data, and vertical inter-layer difference features are integrated.

[0102] For example, when forming meteorological source-scale features, the forecast data of each meteorological source in the key input feature subset are retained, and the wind speed difference features and observation bias features between meteorological sources are combined with the forecast data of each meteorological source to form a meteorological source-scale organization.

[0103] In the above embodiments, the vertical interlayer difference feature is used to characterize the vertical wind profile difference, the wind speed difference feature between meteorological sources is used to reflect the forecast differences between different meteorological sources, and the observation deviation feature is used to reflect the difference between the forecast wind speed and the wind speed measured by the wind tower.

[0104] For example, after the three types of scale features are formed, they are standardized using standard scores to obtain standardized time-scale features, standardized spatial-scale features, and standardized meteorological source-scale features. The three types of standardized scale features are then concatenated along the feature dimensions to form unified multi-scale input data, whose tensor dimensional relationship can be expressed as: ; in, This represents multi-scale input data. Represents the space of real numbers. Indicates the length of the history window. Represents the total feature dimension. This indicates that the multi-scale input data consists of a historical window dimension and a feature dimension.

[0105] In this embodiment, temporal scale features are used to characterize instantaneous changes and trend evolution, spatial scale features are used to characterize differences in data from different vertical layers and vertical wind profiles, and meteorological source scale features are used to characterize differences and complementary information in multi-source forecasts. Based on this, the three types of scale features are standardized and concatenated along the feature dimensions, so that the three types of scale information form a unified input that can be used by both prediction paths, which is beneficial to improving the predictive adaptability when there are differences in multi-source forecasts.

[0106] Figure 8 A flowchart of a local attention-enhanced lightweight linear prediction path provided for some embodiments of this application.

[0107] Reference Figure 8 This includes the following steps: S801 inputs multi-scale input data into parallel channel-independent linear mapping branches, cross-sequence shortcut connection branches, and local attention branches. S802a, the channel-independent linear mapping branch responds to the input of multi-scale input data by using independent weight matrices and biases to map the historical sequence of each feature channel in the multi-scale input data to the prediction space, obtain the mapping result of each feature channel, and average the mapping results to obtain the first branch output; S802b, the cross-sequence shortcut connection branch responds to the input of multi-scale input data. It uses a shared linear layer to directly project the historical sequences of each feature channel in the multi-scale input data to the prediction space, obtains the projection results of each feature channel, and performs mean processing on the projection results to obtain the output of the second branch; S802c, the local attention branch responds to the input of multi-scale input data, performs a linear transformation on the multi-scale input data to obtain a query matrix, a key matrix, and a value matrix. Based on the query matrix, key matrix, and value matrix, the attention range of each time step is limited to a local window with a preset radius and local attention is calculated to obtain the local attention output. Global average pooling is performed on the local attention output to obtain the pooling result, and the pooling result is linearly mapped to the prediction space to obtain the third branch output. S803 adds the outputs of the first branch, the second branch, and the third branch to obtain the branch fusion result; S804 performs fully connected projection and layer normalization on the branch fusion results in sequence to obtain data-driven prediction results.

[0108] For example, in the channel-independent linear mapping branch, each feature channel in the multi-scale input data is assigned an independent weight matrix and bias. The historical sequence of the corresponding feature channel is directly mapped to the prediction space. Then, the mapping results of all feature channels are averaged to obtain the output of the first branch. The calculation process can be represented as follows: ; in, This indicates the output of the first branch. Indicates the number of feature channels. Indicates the feature channel index. Indicates the first The weight matrix corresponding to each feature channel Indicates the first in the multi-scale input data The historical sequence of each feature channel Indicates the first The bias corresponding to each feature channel.

[0109] In some embodiments, the channel-independent linear mapping branch preserves the independent contribution of each feature channel to the prediction result and compresses the multi-channel mapping result into an intermediate representation of univariate prediction.

[0110] For example, in the cross-sequence shortcut connection branch, a linear layer shared by all feature channels is used to directly project the historical sequences of each feature channel in the multi-scale input data into the prediction space. Then, the projection results of all feature channels are averaged to obtain the output of the second branch. ; in, This indicates the output of the second branch. Indicates the number of feature channels. Indicates the feature channel index. This represents the linear layer weights shared by all feature channels. Indicates the first in the multi-scale input data The historical sequence of each feature channel This indicates the bias of the shared linear layer.

[0111] In this embodiment, the cross-sequence shortcut connection branch directly projects the original input sequence into the prediction space to preserve the original temporal dependencies and complements the channel-independent linear mapping branch.

[0112] For example, in the local attention branch, a linear transformation is performed on the multi-scale input data to obtain a query matrix, a key matrix, and a value matrix. Based on this, for each time step, the attention scope is limited to a local window of a preset radius, ensuring that the current time step only interacts with its immediate preceding and following time steps. Further, after completing the local attention calculation based on the query matrix, key matrix, and value matrix, the local attention output is obtained.

[0113] For example, after obtaining the local attention output, global average pooling is performed on the local attention output, and then the pooling result is linearly mapped to the prediction space to obtain the third branch output: ; in, This indicates the output of the third branch. This represents the weight matrix that maps the pooling results to the prediction space. Indicates global average pooling. This refers to local attention computation that limits the scope of attention to a local window. This represents the multi-scale input data that inputs the local attention branch.

[0114] In this embodiment, the local attention branch weights the recent time steps according to the local interaction relationship between neighboring time steps, so that the path can focus on the fluctuation pattern of the neighboring time steps, while reducing the computational overhead of global attention calculation by limiting the attention range.

[0115] As one specific implementation, the overall computational relationship of the local attention-enhanced lightweight linear prediction path is represented as follows: ; in, This indicates data-driven prediction results. Representation layer normalization, This represents the weight matrix of the fully connected projection layer. This indicates the output of the first branch. This indicates the output of the second branch. This indicates the output of the third branch. This indicates the bias of the fully connected projection layer.

[0116] In this embodiment, the channel-independent linear mapping branch, the cross-sequence shortcut connection branch, and the local attention branch extract the independent temporal patterns, original time dependencies, and local dynamic patterns of each feature channel, respectively. The outputs of the three branches are summed, fully connected, projected, and normalized to form a data-driven prediction result. This enables the data-driven prediction process to take into account both global linearity and local fluctuations, and reduces the computational overhead caused by using global attention.

[0117] Figure 9 The flowchart illustrates the prediction strategy based on cascaded short and long windows, provided for some embodiments of this application.

[0118] Reference Figure 9 This includes the following steps: S901a: Extract the short-window power sequence from the historical power sequence, perform first-order linear fitting on the short-window power sequence to obtain the short-window fitting result, and obtain the short-window relative volatility based on the ratio of the residual standard deviation of the short-window power sequence relative to the short-window fitting result to the sum of the absolute value of the mean of the short-window power sequence and the division-to-zero constant. S901b: Perform first-order difference on adjacent power values ​​in the short-window power sequence, and obtain the short-window first-order difference autocorrelation coefficient based on the ratio of the sum of the product of adjacent first-order differences to the sum of the squares of the first-order differences. S902, when the relative volatility of the short window is lower than the preset short window stability threshold and the first-order difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, the long window power sequence is extracted from the historical power sequence, the long window power sequence is fitted with a first-order linear model to obtain the long window fitting result, and the relative volatility of the long window is obtained according to the ratio of the residual standard deviation of the long window power sequence relative to the long window fitting result to the sum of the absolute value of the mean of the long window power sequence and the division-to-zero constant. The first-order difference is performed on the adjacent power values ​​in the long window power sequence, and the first-order difference autocorrelation coefficient of the long window is obtained according to the ratio of the sum of the product of the adjacent first-order differences to the sum of the squares of the first-order differences. S903a: When the relative volatility of the short window is lower than the preset short window stability threshold, the first-order difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, and the relative volatility of the long window is lower than the preset long window stability threshold and the first-order difference autocorrelation coefficient of the long window is higher than the preset oscillation threshold, the current wind condition is determined to be a true trend wind condition, and linear extrapolation is determined as the prediction strategy. S903b: When the relative volatility of the short window is lower than the preset short window stability threshold, the first difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, and the relative volatility of the long window is not lower than the preset long window stability threshold or the first difference autocorrelation coefficient of the long window is not higher than the preset oscillation threshold, the current wind condition is determined to be a pseudo-stable wind condition, and the moving average trend decomposition extrapolation is determined as the prediction strategy. S903c determines the current wind condition as a strong oscillating wind condition when the short-window relative volatility is not lower than the preset short-window stability threshold or the short-window first-order difference autocorrelation coefficient is not higher than the preset oscillation threshold, and determines the conservative median forecast as the forecasting strategy.

[0119] In this embodiment of the application, the hierarchical adaptive prediction path first extracts the short-window power sequence from the historical power sequence, and then determines the short-window relative volatility and the short-window first-order difference autocorrelation coefficient based on the same short-window power sequence.

[0120] For example, only when both short-window indicators meet the corresponding conditions will the long-window power sequence be extracted from the historical power sequence and the two long-window indicators be determined. If the short-window conditions are not met, the long-window processing is not triggered, and the process directly enters the strong oscillation wind condition branch.

[0121] For example, for a power sequence of any length, a first-order linear fit is performed on the power sequence to obtain the corresponding window fitting result. The residual standard deviation of the power sequence relative to the window fitting result is determined, and the relative volatility is obtained based on the ratio of the residual standard deviation to the sum of the absolute value of the mean of the power sequence and the division-by-zero constant. This process can be expressed by the following formula: ; in, Indicates length is The relative volatility of the window power sequence, Indicates the window length. Indicates length is The standard deviation of the residuals of the window power sequence relative to the first-order linear fitting result. Indicates length is The average value of the window power sequence, This represents a constant that prevents division by zero.

[0122] In this embodiment of the application, relative volatility is a dimensionless quantity.

[0123] For example, the smaller the relative volatility, the closer the corresponding window power sequence is to the first-order linear fitting result.

[0124] As a specific implementation method, the first-order difference can be performed on adjacent power values ​​in the same window power sequence, and the first-order difference autocorrelation coefficient can be obtained by comparing the sum of the product of adjacent first-order differences with the sum of the squares of the first-order differences. This process can be expressed by the following formula: ; in, Indicates length is The first-order difference autocorrelation coefficient of the window power sequence, Indicates the window length. Indicates the time index within the window. This represents the time index corresponding to the current moment. Indicates time index The power value at that location, Indicates time index The power value at that location, Indicates time index The power value at that location.

[0125] In this embodiment, the first-order differential autocorrelation coefficient is used to quantify the degree of oscillation reversal between adjacent power changes. The closer the coefficient is to negative one, the more frequently the signs of adjacent changes alternate, and the higher the risk of erroneous extensions when using linear extrapolation.

[0126] For example, when the relative volatility of the short window is lower than the preset short window stability threshold and the first-order difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, the short window power sequence meets the conditions for entering the long window. At this time, the long window power sequence is extracted from the historical power sequence, and the relative volatility and the first-order difference autocorrelation coefficient of the long window are determined in parallel according to the same processing relationship as the short window.

[0127] As a specific judgment scenario, if the short window meets the corresponding conditions, and the relative volatility of the long window is lower than the preset long window stability threshold and the first-order difference autocorrelation coefficient of the long window is higher than the preset oscillation threshold, then the current wind condition is determined to be a true trend wind condition, and a linear extrapolation strategy is determined.

[0128] As another specific judgment scenario, if the short window meets the corresponding conditions, but the relative volatility of the long window is not lower than the preset long window stability threshold or the first-order difference autocorrelation coefficient of the long window is not higher than the preset oscillation threshold, then the current wind condition is determined to be a pseudo-stationary wind condition, and the moving average trend decomposition extrapolation strategy is determined.

[0129] As another specific judgment scenario, if the relative volatility of the short window is not lower than the preset short window stability threshold or the first-order difference autocorrelation coefficient of the short window is not higher than the preset oscillation threshold, then the current wind condition is determined to be a strong oscillating wind condition, and a conservative median forecasting strategy is determined.

[0130] As one specific implementation, a short window stability threshold of 0.1 can be preset, a long window stability threshold of 0.25 can be preset, and an oscillation threshold of -0.5 can be preset.

[0131] In this embodiment, the short window and long window are cascaded to first determine whether the local window has the characteristics of smoothness and weak reversal. Then, when the short window meets the conditions, the long window is used to examine the fluctuation and reversal relationship over a longer time range. This can distinguish between true trend conditions, pseudo-stable conditions where the local window is smooth but the long window has fluctuations or reversals, and strong oscillation conditions where the short window has strong oscillations. This reduces the possibility of misjudging local smoothness as a continuous trend.

[0132] Figure 10 This is a flowchart illustrating the determination of time-series predicted values ​​based on a prediction strategy, provided for some embodiments of this application.

[0133] Reference Figure 10 This includes the following steps: S1001a, when the prediction strategy is linear extrapolation, the power values ​​of adjacent time steps in the short window power sequence are subtracted to obtain the adjacent power difference. The adjacent power differences are weighted and summed to obtain the weighted slope. The power at the current moment of the historical power sequence is taken as the starting point, and linear extrapolation is performed according to the weighted slope and the trend strength coefficient to obtain the first time series prediction value. The first time series prediction value is determined as the time series prediction value. S1001b, when the prediction strategy is moving average trend decomposition extrapolation, the long window power series is subjected to moving average to obtain the trend component series. The trend component series is extrapolated by univariate linear regression with time index as independent variable to obtain the second time series prediction value. The second time series prediction value is determined as the time series prediction value. S1001c, when the prediction strategy is conservative median prediction, the most recent window power sequence is extracted from the historical power sequence, the median of the most recent window power sequence is determined, the third time series prediction value is obtained, and the third time series prediction value is determined as the time series prediction value.

[0134] S1002, the time series prediction value is fused with the deviation correction reference value to obtain the physical rule prediction result.

[0135] In this embodiment, S1001a, S1001b and S1001c are three branches, and at the same prediction time, only the corresponding branch is executed according to the currently determined prediction strategy.

[0136] For example, in the linear extrapolation branch, the power values ​​of adjacent time steps in the short window power sequence are subtracted to obtain the adjacent power difference, and the adjacent power differences are weighted and summed to obtain the weighted slope.

[0137] As one specific implementation, the power difference between three adjacent values ​​within the short window is calculated and assigned a corresponding weight. The weighted slope is determined according to the following relationship: ; in, Indicates the weighted slope. Indicates the first The weights corresponding to the adjacent power differences Indicates the first The power difference between adjacent values, An index representing the difference between adjacent power values.

[0138] Starting with the current power in the historical power sequence, a linear extrapolation is performed based on the weighted slope, trend strength coefficient, and the time step to be predicted: ; in, Indicates the number of times after the current moment. The first time-series prediction value at each time step. Indicates the power at the current moment. This represents the time index corresponding to the current moment. Indicates the time step to be predicted. Indicates the trend strength coefficient. Indicates the weighted slope. This indicates the true trend and risk conditions corresponding to linear extrapolation.

[0139] For example, in the moving average trend decomposition extrapolation branch, a moving average is applied to the long-window power series to obtain the trend component series: ; in, , Indicates time index The trend component at the location, Indicates the length of the long window. Indicates the index of the power value within the long window. Indicates time index The power value at that location, This indicates the time index corresponding to the trend component. This represents the time index corresponding to the current moment.

[0140] In this embodiment, the moving average is used to filter out high-frequency oscillations in the long-window power sequence, enabling univariate linear regression to extrapolate based on the dominant evolution direction in the long window.

[0141] For example, in the conservative median prediction branch, the most recent window power series is extracted from the historical power series, and the median of the most recent window power series is used as the third time series prediction value: ; in, This represents the third time series prediction value. This represents the base forecast value formed by the conservative median forecast. This represents median operations. Indicates the length of the most recent window. This represents the time index corresponding to the current moment. to This represents the historical power values ​​in the most recent window power sequence.

[0142] In this embodiment, the median of the power sequence of the most recent window is used so that the time series prediction value can fall within the range of the historical power value of the most recent window, and the deviation amplification caused by continuing to extrapolate when the actual change direction is opposite to the recent trend is reduced.

[0143] In this embodiment, linear extrapolation, moving average trend decomposition extrapolation, and conservative median prediction are triggered according to true trend wind conditions, pseudo-stationary wind conditions, and strong oscillation wind conditions, respectively, so that the time series prediction process can be adapted to the historical power states of smooth trends, local smoothing but overall turning points, and short-window strong oscillations, respectively.

[0144] Figure 11 This is a flowchart illustrating the fusion of timing prediction values ​​and deviation correction reference values ​​provided in some embodiments of this application.

[0145] Reference Figure 11 This includes the following steps: S1101a, Determine the ratio of the deviation correction reference value to the rated capacity of the unit; S1101b, Determine the deviation rate of the time series prediction value relative to the deviation correction reference value; S1102, apply soft constraints to the deviation rate according to the ratio, and perform a weighted average of the time series prediction value and the deviation correction reference value under the soft constraints to obtain the physical rule prediction result.

[0146] In some embodiments, the calculation relationship of the deviation rate can be expressed as: ; Among them, Indicates the deviation rate. This represents the time-series forecast value generated by the current forecasting strategy. This represents the deviation correction reference value determined based on historical forecast deviation statistics. This indicates the time step to be predicted.

[0147] As a feasible implementation method, the deviation rate is limited when the deviation correction reference value corresponds to a high-power region, and the weight of the deviation correction reference value in the weighted average is reduced when the deviation correction reference value corresponds to a low-power region. Therefore, by adjusting the degree of deviation between the time-series predicted value and the deviation correction reference value according to the power level at which the deviation correction reference value is located, the amplification of deviations in the time-series extrapolation results can be suppressed.

[0148] Figure 12The diagram shows the structure of a dual-path parallel fusion wind power ultra-short-term power prediction system provided in some embodiments of this application.

[0149] Reference Figure 12 The wind power ultra-short-term power prediction system includes a data construction module 1201, a dual-path prediction module 1202, and a prediction output module 1203.

[0150] The data construction module 1201 is used to acquire wind farm operation data and multi-source numerical weather forecast data to construct multi-scale input data; it constructs a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path for parallel processing; wherein, the local attention-enhanced lightweight linear prediction path is trained based on the predicted power sequence and the real power sequence, and the predicted power sequence and the real power sequence are constrained by the sequence shape preservation loss based on the time domain local statistics and the frequency domain amplitude spectrum. The dual-path prediction module 1202 is used to input multi-scale input data into a local attention-enhanced lightweight linear prediction path to obtain data-driven prediction results; based on the historical power sequence in the multi-scale input data, a prediction strategy is determined, and the time-series prediction value is fused with the deviation correction reference value to obtain the physical rule prediction result; wherein, the time-series prediction value is determined based on the prediction strategy, and the deviation correction reference value is determined based on the historical prediction deviation statistics. The prediction output module 1203 is used to perform ensemble prediction on the data-driven prediction results and the physical rule prediction results to obtain the wind power ultra-short-term power prediction sequence.

[0151] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0152] Although the operations are described in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the operations shown to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0153] The methods and systems of this application can be implemented using standard programming techniques, utilizing rule-based logic or other logic to implement various method steps. It should also be noted that the terms "system" and "module" as used herein and in the claims are intended to include implementations using one or more lines of software code and / or hardware implementations and / or devices for receiving input.

[0154] Any step, operation, or procedure described herein may be performed or implemented using one or more hardware or software modules, either alone or in combination with other devices. In one embodiment, the software module is implemented using a computer program product comprising a computer-readable medium containing computer program code, which is executable by a computer processor to perform any or all of the described steps, operations, or procedures.

[0155] The foregoing description of implementations of this application has been provided for illustrative and descriptive purposes. The foregoing description is not exhaustive and is not intended to limit this application to the exact forms disclosed. Various modifications and variations may exist in accordance with the foregoing teachings, or may arise from practice of this application. These embodiments were chosen and described to illustrate the principles of this application and its practical application, enabling those skilled in the art to utilize this application in various implementations and modifications to suit the specific purpose of the concept.

[0156] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0157] It can be further understood that, unless otherwise specified, "connection" includes both direct connections where no other components exist between the two parties and indirect connections where other components exist between them.

[0158] It is further understood that although the operations are described in a specific order in the accompanying drawings in the embodiments of this application, this should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all the operations shown to be performed to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0159] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the field of this application that are not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0160] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

[0161] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting ultra-short-term wind power using a dual-path parallel fusion approach, characterized in that, include: Acquire wind farm operation data and multi-source numerical weather prediction data to construct multi-scale input data; Construct a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path for parallel processing; wherein the local attention-enhanced lightweight linear prediction path is trained based on a predicted power sequence and a true power sequence, and the predicted power sequence and the true power sequence are constrained by a sequence shape preservation loss based on temporal local statistics and frequency domain amplitude spectrum. The multi-scale input data is input into the local attention-enhanced lightweight linear prediction path to obtain data-driven prediction results. Based on the historical power sequence in the multi-scale input data, a prediction strategy is determined, and the time-series predicted value and the deviation correction reference value are fused based on the hierarchical adaptive prediction path to obtain the physical rule prediction result; wherein, the time-series predicted value is determined based on the prediction strategy, and the deviation correction reference value is determined based on the historical prediction deviation statistics. The data-driven prediction results and the physical rule prediction results are combined for prediction to obtain the wind power ultra-short-term power prediction sequence.

2. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 1, characterized in that, The acquisition of wind farm operation data and multi-source numerical weather prediction data to construct multi-scale input data includes: Acquire the wind farm operation data and the multi-source numerical weather forecast data; The wind farm operation data and the multi-source numerical weather forecast data are aligned in time dimension, anomaly sample processing and normalization processing are performed to obtain basic feature data. Based on the basic feature data, derived features are constructed, the basic feature data and the derived features are combined to obtain a candidate feature set, and a subset of key input features is selected from the candidate feature set; The key input feature subset is organized into multiple scales to obtain the multi-scale input data.

3. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 2, characterized in that, The wind farm operation data includes actual grid-connected power, available power, nacelle wind speed, and wind speed, wind direction, temperature, and relative humidity at each vertical level of the meteorological tower. The multi-source numerical weather forecast data includes numerical weather forecast wind speed, wind direction, temperature, humidity, and precipitation from different meteorological sources at different vertical levels. The process of aligning the wind farm operation data and the multi-source numerical weather forecast data by time dimension, processing outlier samples, and normalizing them yields basic feature data, including: The wind farm operation data and the multi-source numerical weather forecast data are inner-joined and aligned to obtain equally spaced time series samples. The equally spaced time series samples are divided into training data, validation data, and test data; Anomaly detection is performed on each continuous feature in the training data using a statistical threshold based on the interquartile range to obtain the first candidate anomaly sample. The continuous features in the training data are standardized to obtain the anomaly detection feature space. K-means clustering and isolated forest are used to judge anomalies in the anomaly detection feature space to obtain the second candidate anomaly sample and the third candidate anomaly sample. The intersection of the first candidate abnormal sample, the second candidate abnormal sample, and the third candidate abnormal sample is removed from the training data to obtain the cleaned training data; Based on the cleaned training data, the minimum and maximum values ​​of each feature are determined, and the cleaned training data, the validation data, and the test data are normalized based on the minimum and maximum values ​​to obtain the basic feature data.

4. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 2, characterized in that, The step of selecting a subset of key input features from the candidate feature set includes: The candidate feature set is input into the random forest model and the extreme gradient boosting model respectively to obtain the random forest prediction result and the extreme gradient boosting prediction result. Based on the average reduction in Gini impurity caused by the splitting of each decision tree node in the random forest model, the first feature importance of each candidate feature is obtained. The second feature importance of each candidate feature is obtained based on the average gain contribution of each split point in the extreme gradient boosting model. The random forest prediction results or the extreme gradient boosting prediction results are subjected to Shapley additive interpretation to obtain the third feature importance of each candidate feature; The importance of the first feature, the second feature, and the third feature of each candidate feature are normalized and then weighted and summed to obtain the comprehensive score of each candidate feature. The candidate features are sorted from high to low according to the comprehensive score, and the candidate features are selected in turn until the cumulative score of the selected candidate features exceeds a preset threshold, thus obtaining the key input feature subset.

5. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 4, characterized in that, The process of organizing the key input feature subset into multi-scale data to obtain the multi-scale input data includes: Extract current time features, lag features, and sliding statistical features with the same timestamp from the key input feature subset, and superimpose the current time features, the lag features, and the sliding statistical features to obtain time scale features; Extract the layered data of the wind measurement tower, the cabin height data, and the multi-vertical layer data of numerical weather prediction from the key input feature subset to determine the vertical inter-layer difference features. Then, integrate the layered data of the wind measurement tower, the cabin height data, the multi-vertical layer data of numerical weather prediction, and the vertical inter-layer difference features to obtain the spatial scale features. The forecast data of each meteorological source in the key input feature subset, the wind speed difference feature between the meteorological sources, and the observation deviation feature are organized at the meteorological source scale to obtain meteorological source scale features; Standardized temporal scale features, spatial scale features, and meteorological source scale features are obtained by standardizing the time scale features, spatial scale features, and meteorological source scale features using standard scores. The standardized time-scale features, the standardized spatial-scale features, and the standardized meteorological source-scale features are concatenated along the feature dimensions to obtain the multi-scale input data.

6. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 1, characterized in that, The step of inputting the multi-scale input data into the locally attention-enhanced lightweight linear prediction path to obtain data-driven prediction results includes: The multi-scale input data is respectively input into the parallel channel-independent linear mapping branch, the cross-sequence shortcut connection branch, and the local attention branch; The channel-independent linear mapping branch responds to the input of the multi-scale input data by using independent weight matrices and biases to map the historical sequence of each feature channel in the multi-scale input data to the prediction space, thereby obtaining the mapping result of each feature channel. The mapping results are then averaged to obtain the first branch output. The cross-sequence shortcut connection branch responds to the input of the multi-scale input data by using a shared linear layer to directly project the historical sequences of each feature channel in the multi-scale input data to the prediction space, thereby obtaining the projection results of each feature channel. The projection results are then averaged to obtain the output of the second branch. The local attention branch responds to the input of the multi-scale input data by performing a linear transformation on the multi-scale input data to obtain a query matrix, a key matrix, and a value matrix. Based on the query matrix, the key matrix, and the value matrix, the attention range of each time step is limited to a local window with a preset radius and local attention is calculated to obtain a local attention output. The local attention output is then subjected to global average pooling to obtain a pooling result, and the pooling result is linearly mapped to the prediction space to obtain the third branch output. The first branch output, the second branch output, and the third branch output are added together to obtain the branch fusion result; The branch fusion results are sequentially subjected to fully connected projection and layer normalization to obtain the data-driven prediction results.

7. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 1, characterized in that, The step of determining the prediction strategy based on the historical power sequence in the multi-scale input data includes: A short-window power sequence is extracted from the historical power sequence, and a first-order linear fit is performed on the short-window power sequence to obtain the short-window fitting result. The short-window relative volatility is obtained based on the ratio of the residual standard deviation of the short-window power sequence relative to the short-window fitting result to the sum of the absolute value of the mean of the short-window power sequence and the division-to-zero constant. The first-order difference of adjacent power values ​​in the short-window power sequence is performed, and the autocorrelation coefficient of the first-order difference of the short window is obtained by the ratio of the product of the adjacent first-order differences to the square of the first-order differences. When the relative volatility of the short window is lower than a preset short window stability threshold and the first-order difference autocorrelation coefficient of the short window is higher than a preset oscillation threshold, a long window power sequence is extracted from the historical power sequence. The long window power sequence is then subjected to first-order linear fitting to obtain a long window fitting result. The relative volatility of the long window is obtained by the ratio of the residual standard deviation of the long window power sequence relative to the long window fitting result to the sum of the absolute value of the mean of the long window power sequence and the division-to-zero constant. The first-order difference of adjacent power values ​​in the long window power sequence is then performed. The first-order difference autocorrelation coefficient of the long window is obtained by the ratio of the sum of the product of adjacent first-order differences to the sum of the squares of the first-order differences. When the relative volatility of the short window is lower than the preset short window stability threshold, the first-order difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, and the relative volatility of the long window is lower than the preset long window stability threshold and the first-order difference autocorrelation coefficient of the long window is higher than the preset oscillation threshold, the current wind condition is determined to be a true trend wind condition, and linear extrapolation is determined as the prediction strategy. When the relative volatility of the short window is lower than the preset short window stability threshold, the first difference autocorrelation coefficient of the short window is higher than the preset oscillation threshold, and the relative volatility of the long window is not lower than the preset long window stability threshold or the first difference autocorrelation coefficient of the long window is not higher than the preset oscillation threshold, the current wind condition is determined to be a pseudo-stable wind condition, and the moving average trend decomposition extrapolation is determined as the prediction strategy. When the relative volatility of the short window is not lower than the preset short window stability threshold or the first-order difference autocorrelation coefficient of the short window is not higher than the preset oscillation threshold, the current wind condition is determined to be a strong oscillating wind condition, and the conservative median prediction is determined as the prediction strategy.

8. The wind power ultra-short-term power prediction method with dual-path parallel fusion according to claim 7, characterized in that, The time-series prediction values ​​are determined in the following manner: When the prediction strategy is linear extrapolation, the power values ​​of adjacent time steps in the short window power sequence are subtracted to obtain the adjacent power difference. The adjacent power differences are weighted and summed to obtain the weighted slope. Starting from the current power of the historical power sequence, linear extrapolation is performed according to the weighted slope and the trend strength coefficient to obtain the first time series prediction value. The first time series prediction value is determined as the time series prediction value. When the prediction strategy is moving average trend decomposition extrapolation, the long window power sequence is subjected to moving average to obtain trend component sequence. The trend component sequence is extrapolated by univariate linear regression with time index as independent variable to obtain second time series prediction value. The second time series prediction value is determined as the time series prediction value. When the prediction strategy is conservative median prediction, the most recent window power sequence is extracted from the historical power sequence, the median of the most recent window power sequence is determined, a third time series prediction value is obtained, and the third time series prediction value is determined as the time series prediction value.

9. The dual-path parallel fusion wind power ultra-short-term power prediction method according to claim 1 or 7, characterized in that, The process of fusing time-series predicted values ​​with bias correction reference values ​​to obtain physical rule prediction results includes: Determine the ratio of the deviation correction reference value to the unit's rated capacity; Determine the deviation rate of the time-series predicted value relative to the deviation correction reference value; A soft constraint is applied to the deviation rate according to the ratio, and a weighted average is performed on the time-series predicted value and the deviation correction reference value under the soft constraint to obtain the physical rule prediction result.

10. A dual-path parallel fusion wind power ultra-short-term power prediction system, characterized in that, include: The data construction module is used to acquire wind farm operation data and multi-source numerical weather forecast data to construct multi-scale input data; Construct a local attention-enhanced lightweight linear prediction path and a hierarchical adaptive prediction path for parallel processing; wherein the local attention-enhanced lightweight linear prediction path is trained based on a predicted power sequence and a true power sequence, and the predicted power sequence and the true power sequence are constrained by a sequence shape preservation loss based on temporal local statistics and frequency domain amplitude spectrum. A dual-path prediction module is used to input the multi-scale input data into the local attention-enhanced lightweight linear prediction path to obtain a data-driven prediction result; based on the historical power sequence in the multi-scale input data, a prediction strategy is determined, and the time-series prediction value is fused with the deviation correction reference value to obtain a physical rule prediction result; wherein, the time-series prediction value is determined based on the prediction strategy, and the deviation correction reference value is determined based on the historical prediction deviation statistics. The prediction output module is used to perform a combined prediction on the data-driven prediction results and the physical rule prediction results to obtain the wind power ultra-short-term power prediction sequence.