A day-ahead wind power ramp event prediction method

Abstract

The application discloses a kind of day-ahead wind power power ramp event prediction methods, comprising the following steps: S1.day-ahead wind power prediction: using extreme value driving model, based on historical wind power and meteorological characteristics, produce the wind power prediction of Q time points of next day;S2.predicting confidence interval construction: using wind power oriented conformal inference method, for each future time point, construct confidence interval C;S3.ramp event detection based on confidence interval: the prediction result of next day wind power ramp event is obtained by confidence perception detection algorithm.The application improves the prediction accuracy and reliability of wind power ramp event with significant operation risk by fusing multi-scale time series feature analysis, customized loss function for extreme event and adaptive confidence interval construction technology with statistical guarantee.

Description

Technical Field

[0001] This invention relates to the field of wind power grid connection and power system dispatching technology, and in particular to a method for predicting day-ahead wind power ramp-up events. Background Technology

[0002] With the acceleration of the global energy transition, the proportion of wind power in the power structure continues to rise. However, the inherent intermittency and volatility of wind power, especially wind power ramp-up events (WPREs), which are large and rapid changes in wind farm output power within a short period of time (such as changes exceeding 20% ​​of installed capacity within two hours), pose a severe challenge to the stability of the power system, dispatch efficiency, and market operation. Sudden upward changes may lead to wind curtailment, while sudden downward changes may force the use of expensive reserve resources and even trigger power supply risks. Therefore, achieving high-precision, high-confidence day-ahead WPRE forecasts is of vital importance for forward-looking grid dispatch, rational allocation of reserve capacity, and ensuring the safe and economical operation of the grid.

[0003] Currently, prediction methods for WPREs can be mainly divided into two categories: direct methods and indirect methods. Direct methods train classification or regression models based on historical WPRE event labels. Although the objective is clear, the sparse nature of WPRE event records results in sparse and imbalanced label data available for training, leading to poor model generalization ability.

[0004] The indirect method first predicts the complete day-ahead wind power time series, and then identifies WPREs from the predicted curve using a specific abrupt change detection algorithm, effectively avoiding the problem of data sparsity. However, the drawback of this type of method is that existing wind power prediction models are difficult to accurately capture extreme power, and the identification of WPREs is highly dependent on the prediction accuracy of extreme power.

[0005] In summary, although wind power prediction technology has made great strides, existing technologies have obvious shortcomings when facing the specific and critical engineering problem of WPREs: direct methods are constrained by data bottlenecks, while indirect methods are limited in effectiveness due to the mismatch between the target of the core prediction model and the physical characteristics of WPREs. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting day-ahead wind power ramping events. By integrating multi-scale time series feature analysis, customized loss functions for extreme events, and adaptive confidence interval construction technology with statistical guarantees, the method improves the prediction accuracy and reliability of wind power ramping events with significant operational risks.

[0007] The objective of this invention is achieved through the following technical solution: a method for predicting day-ahead wind power ramp-up events, comprising the following steps:

[0008] S1. Day-ahead wind power forecast: Using an extreme value driven model, based on historical wind power and meteorological characteristics, wind power forecasts are generated for Q time points on the next day;

[0009] S2. Construction of Confidence Intervals for Prediction: Using the conformal inference method guided by wind power, a confidence interval C is constructed for each future time point;

[0010] S3. Confidence Interval-Based Ramp-up Event Detection: The prediction results of the wind power ramp-up event for the next day are obtained through a confidence-aware detection algorithm.

[0011] The beneficial effects of this invention are as follows: First, this invention designs an extreme value-driven learning model. By introducing a hybrid frequency feature enhancement module, it simultaneously captures the long-term and short-term time-series dynamics of wind power. It also employs a customized extreme loss function, shifting the model's learning focus from optimizing average accuracy to accurately predicting the power extremes that determine sudden change events, thus overcoming the problem of missed sudden change reports caused by the "smoothing" of traditional prediction models. Second, this invention pioneers a wind power-oriented conformal inference method, capable of generating prediction confidence intervals with strict statistical guarantees and adaptively varying widths, exhibiting significantly better coverage reliability than traditional methods. Finally, based on this confidence interval, this invention proposes a confidence-aware sudden change detection algorithm. By dynamically adjusting the detection threshold, it achieves personalized and risk-controllable sudden change event identification. Ultimately, while ensuring high accuracy, it greatly improves the accuracy of day-ahead wind power sudden change event prediction, with negligible computational overhead, combining high innovation with engineering practicality. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0014] like Figure 1 As shown, a method for predicting day-ahead wind power ramp-up events includes the following steps:

[0015] S1. Day-ahead wind power forecast: Using an extreme value driven model, based on historical wind power and meteorological characteristics, wind power forecasts are generated for Q time points on the next day;

[0016] S2. Construction of Confidence Intervals for Prediction: Using the conformal inference method guided by wind power, a confidence interval C is constructed for each future time point;

[0017] S3. Confidence Interval-Based Ramp-up Event Detection: The prediction results of the wind power ramp-up event for the next day are obtained through a confidence-aware detection algorithm.

[0018] Considering that wind power forecasts are based on a "daily" timescale, while WPREs (Wind Power Reduction Events) represent "short-term, dramatic changes within hours" (e.g., a 20% change in installed capacity within two hours), the forecasting model must simultaneously capture both long-term daily trends and short-term, hourly abrupt changes. This is extremely challenging for traditional time series models. Furthermore, the identification of WPREs is highly dependent on the maximum and minimum power points. Therefore, the model must not only improve overall average forecast accuracy but also accurately predict extreme wind power values. Existing methods often optimize the average error, thereby masking key points of dramatic change. To address this, we propose the solution outlined in this application.

[0019] In step S1, the day-ahead wind power forecast includes:

[0020] S101. Construct an extremum-driven model, including a feature embedding module, an encoder, and a projection layer;

[0021] The feature embedding module is used to convert historical data into feature representations suitable for deep learning models, including variational mode decomposition units, embedding networks, and embedding units.

[0022] The encoder contains an L-layer structure, with each layer encoded by a HFFE module. It uses an attention mechanism to extract short-term and long-term dependency structures. The HFFE module refers to the Hybrid Frequency Feature Enhancement Module.

[0023] The projection layer decodes the encoder output into a daily wind power prediction.

[0024] S102. Collect historical data of the target wind farm and construct a database of the target wind farm, including wind power data of N wind turbines in the wind farm at the past K time points, meteorological characteristics W composed of G types of weather in the wind farm, and spatial adjacency matrix A of spatial correlation between N wind turbines.

[0025] S103. For any wind turbine in the wind farm, construct a training dataset:

[0026] (1) From the wind power data of the wind turbine at the past K time points, select wind power data for a consecutive M days. Let the number of time points in each day be Q. The wind power data of the first M-1 days in the consecutive M days constitute the power sequence X(t) of the current wind turbine, forming the input sample. The actual wind power data of the Q time points in the Mth day are used as labels. The features of the input sample and the corresponding labels constitute a training sample, where M>2 and M*Q <K;

[0027] (2) Repeat (1) to obtain multiple training samples, which constitute the training sample set of the current wind turbine;

[0028] S104. Train the extreme value driving model of any wind turbine in the wind farm to obtain the trained extreme value driving model corresponding to the wind turbine.

[0029] S105. For each wind turbine in the wind farm, repeat steps S103 to S104 to obtain the trained extreme value driving model corresponding to each wind turbine in the wind farm.

[0030] S106. For each wind turbine in the wind farm, construct a power sequence using the wind power data from the most recent M-1 days, input it into the trained extreme value driving model corresponding to the wind turbine, and obtain the wind power prediction for Q time points on the next day.

[0031] Step S104 includes:

[0032] For any wind turbine in a wind farm, based on each sample in the training sample set of that wind turbine, the power sequence X(t) is input into the feature embedding module for processing, and the historical data is converted into a feature representation suitable for deep learning model processing, thus forming the embedded feature.

[0033] The output of the feature embedding module is encoded by the encoder, and then the encoded result is sent to the projection layer. In the projection layer, the power prediction values ​​for the next Q time points are obtained by decoding through a multilayer perceptron.

[0034] The loss function is calculated using the power predictions for the next Q time points and the labels of the current samples, and the extreme value-driven model is updated accordingly.

[0035] Once training has been completed for each sample in the training sample set, or the loss function is less than a set threshold, the resulting extreme value-driven model is used as the extreme value-driven model for the current wind turbine.

[0036] Further, the power sequence X(t) is input into the feature embedding module for processing, converting historical data into feature representations suitable for deep learning models, and forming embedded features, including:

[0037] Capturing multi-scale time series features: Using variational mode decomposition units, the power sequence X(t) of each wind turbine is decomposed into M intrinsic IMF components, which constitute IMF features;

[0038] Wind power exhibits high uncertainty and dramatic fluctuations. Without preprocessing, models may only capture the overall trend while ignoring short-term abrupt changes. Therefore, this method introduces Variational Mode Decomposition (VMD) to extract intrinsic oscillation modes at different time scales. VMD decomposes the wind power signal into a set of Intrinsic Mode Functions (IMFs), with each IMF corresponding to a center frequency.

[0039] The decomposed IMF features, spatial adjacency matrix A, meteorological features W, and power sequence X(t) are fed into the embedding network and mapped to a unified D-dimensional feature space to obtain the mapping result of the IMF features. Mapping results of spatial adjacency matrix Mapping results of meteorological characteristics , and the mapping results of the power sequence;

[0040]

[0041] Spatiotemporal embedding is generated through embedding units. Space-meteorological embedding :

[0042] ;

[0043] ;

[0044] This yields the output of the embedded unit. .

[0045] The process of encoding the output of the feature embedding module using an encoder includes:

[0046] The encoder contains an L-layer structure. Each layer encodes through an HFFE block. A feedforward network and normalization processing are added to each HFFE block. The fused features are then passed through another feedforward network and normalization layer to complete the processing of an HFFE block.

[0047] Each layer's HFFE block aims to capture both long and short-term dependencies simultaneously, that is, to capture short-term changes and long-term advantages through frequency attention and feature attention. The processing procedure for the HFFE block in layer l is as follows:

[0048] While VMD has enhanced its ability to capture short-term changes, the long-term trend of wind power is often determined by meteorological characteristics. Therefore, the HFFE block extracts short-term high-frequency characteristics (from IMF) and long-term low-frequency trends (from meteorological characteristics) through two attention mechanisms, and achieves adaptive fusion through a gating mechanism.

[0049] For spatiotemporal embedding By calculating its attention weights on different IMF frequency components through frequency attention, it focuses on short-term, sharp fluctuations associated with sudden changes:

[0050] ;

[0051] in, This represents the hidden state of the l-th layer. This represents the frequency attention calculation result of the l-th layer;

[0052] For space-weather embedding The attention weights of the impact of different meteorological characteristics on power are calculated to provide long-term trend information.

[0053] ;

[0054] in, This represents the result of the feature attention calculation for the l-th layer;

[0055] The two attention outputs are adaptively fused using relevant decision gates. :

[0056] ;

[0057] in, , As a weight for relevant decision-makers, The bias of the relevant decision-makers , This represents element-wise multiplication;

[0058] The hidden state of the encoder's first layer is the mapping result of the power sequence. From the second layer to the (L-1)th layer, each layer obtains... After passing through the feedforward network and the normalization layer, it serves as the hidden state for the next layer, resulting in the Lth layer. The output of the encoder is obtained by passing through a feedforward network and a normalization layer.

[0059] Traditional loss methods (MSE / MAE) cause the model to ignore the most critical extreme wind power points in WPREs. Extreme value driven models enhance the prediction accuracy of extreme points through extreme value loss, the loss function of which is defined as:

[0060] The extreme value-driven model enhances the prediction accuracy of extreme points through extreme value loss, the loss function of which is defined as:

[0061] ;

[0062] The prediction time range is divided into multiple time windows. This represents the set of time points representing extreme values ​​within each time window. This indicates the number of extreme points; the extreme values ​​refer to the maximum and minimum values ​​within a time window, and the time points of the extreme values ​​refer to the times when the extreme values ​​occur within each time window. The weights of the extreme value terms; where Q represents the number of time points within the predicted time range; , This represents the actual and predicted wind power output at time point t.

[0063] The Adam optimizer or stochastic gradient descent implementation used when updating the extreme value driven model.

[0064] While extreme value-driven learning models significantly improve the accuracy of wind power forecasting (especially for extreme values), relying solely on data-driven methods still makes it difficult to directly achieve high-confidence WPRE (Wind Power Revenue Scale) predictions. From an engineering perspective, grid dispatch requires prediction results with clearly defined uncertainties to support risk-controlled operational decisions.

[0065] To address this, the present invention proposes a wind power-oriented conformal inference method for constructing an adaptively variable wind power confidence interval. This allows the model to dynamically adjust the interval width across different time periods while ensuring strict statistical coverage. Specifically, step S2 includes:

[0066] Step S2 includes:

[0067] S201. Based on the extreme value driving model of each wind turbine, predict the wind power of the wind turbine in the most recent two days and the next day; at each predicted time point, superimpose the predicted power of each wind turbine in the wind farm to obtain the predicted wind power value of the wind farm in the most recent two days and the next day.

[0068] By superimposing the actual power of each wind turbine in the wind farm at each point in time within the most recent two days in history, the true value of wind power at each point in time within the most recent two days in history is obtained.

[0069] The prediction period is three days, with Q time points per day, for a total of 3Q time points. The most recent two historical time points are denoted as follows: ~ Record the time of the next day as ~ ;

[0070] Construct a calibration set, including:

[0071] Forecast data This indicates that within the prediction interval, the wind farm starts from the first time point. By the 2Qth time point The set of predicted wind power values;

[0072] Actual data This indicates that within the prediction interval, the wind farm starts from the first time point. By the 2Qth time point The set of actual wind power values;

[0073] S202. Calculate the absolute value of the prediction residual for each time point in the calibration set prediction interval, as the conformal fraction:

[0074] ;

[0075] in, Let represent the conformal fraction at the t-th time point in the prediction interval. , This represents the actual wind power output and the predicted wind power output at time point t.

[0076] S203. Conformal fractions in the calibration set Apply time decay weights :

[0077] ;

[0078] For future time points If i is not greater than Q, use the conformal fraction of the history window. Calculate the weighted 1-α quantile:

[0079] ;

[0080] in, This indicates the time point from the i-th time in the prediction interval. up to the (Q+i)th time point The set of conformal fractions; Indicates to The calculated 1-α quantile; α represents the pre-given confidence level;

[0081] right Select conformal fractions in ascending order until the normalized weights corresponding to the selected conformal fractions are reached. The sum is greater than or equal to 1-α, for the selected conformal fraction, and its weight Multiplication Set to 1; for unselected conformal fractions, multiply by their weights. Set to 0;

[0082] Used to handle extreme cases where the fractional digit condition cannot be met, when After selecting all conformal fractions, normalize the weights. If the sum is still less than 1-α, then It is 1 if it is true, otherwise it is 0;

[0083] in,

[0084] ;

[0085] ;

[0086] S203. Construct confidence intervals:

[0087] future time points Construct upper and lower bounds to obtain the confidence interval C:

[0088] ;

[0089] The upper bound is defined as "+", and the lower bound as "-". This represents the wind power data at the 2Q+i time point predicted using the extreme value driven model;

[0090] Simultaneously, physical feasibility constraints are incorporated:

[0091] ;

[0092] That is, the lower world Cannot be lower than 0, upper bound The power output must not exceed the feasible power range of the wind turbine. ;

[0093] The upper and lower bounds are calculated on a rolling basis as time changes.

[0094] In an embodiment of this application, after completing the interval for the third day, the calibration set is rolled backward: the interval for the fourth day is constructed using the data from the second and third days.

[0095] This process continues until confidence interval predictions for all days are completed. In the embodiments of this application, the proposed wind power-oriented conformal inference method satisfies the following coverage guarantee:

[0096] ;

[0097] That is, the actual coverage rate is greater than or equal to the target coverage rate. ;

[0098] This invention proposes a confidence-aware ramp event detection algorithm that uses the confidence interval of day-ahead wind power prediction to identify wind power ramp events (WPREs) under the premise of controllable uncertainty.

[0099] The core idea of ​​the detection algorithm is:

[0100] The maximum difference between the upper and lower bounds of the confidence interval is compared with a dynamic threshold to determine whether WPREs have occurred.

[0101] In particular, an adaptive ramp threshold is introduced to reduce false positives under the traditional fixed threshold and improve the reliability of ramp detection.

[0102] The design philosophy of adaptive thresholds: Traditional WPRE detection typically uses a fixed threshold (such as 20% of installed capacity). However, in actual forecasting: if the uncertainty of wind power forecasting is large, a fixed threshold may lead to a large number of false alarms; if the forecasting uncertainty is small, a fixed threshold may lead to false negatives.

[0103] Therefore, specifically as in step S3, it includes:

[0104] S301. Define dynamic threshold:

[0105] Adaptive threshold To reduce false alarms, the threshold for each sliding window of two hours is defined as follows:

[0106] ;

[0107] in, Based on a fixed threshold; Indicates the width of the confidence interval at time point t; This indicates the average width of the interval within the window. This represents the average interval width over a day, specifically within Q time points. This is the starting point of the sliding window. The length of the sliding window is set to two hours.

[0108] S302. Using confidence intervals for hill climb detection:

[0109] Slide one up the time series of the predicted day For each time window of a given length, calculate the maximum value of the upper bound of its built-in confidence interval. minimum value of the lower bound :

[0110] ;

[0111] ;

[0112] in, It is the set of upper and lower bounds of the confidence interval for each time point within the time window;

[0113] If satisfied If a wind power ramp-up event occurs in the current window, then determine that a wind power ramp-up event has occurred and output the result. Otherwise, output 0 and continue sliding the window.

[0114] In summary, this invention proposes an extreme value-driven learning model to capture large variations in wind power output and significantly improve the prediction accuracy of extreme values, thereby overcoming the underestimation problem. Specifically:

[0115] - Hybrid Frequency Feature Enhancement Module (HFFE): Used to focus on fine-grained frequency components, enabling the model to retain the abrupt jumps in wind power changes and more accurately reconstruct extreme climbing patterns;

[0116] - Incorporate meteorological data (wind speed, wind direction, temperature, etc.) to improve the model's ability to identify precursors of sudden changes in wind power and enhance the physical interpretability of the predictions;

[0117] - Customized Extreme-based Loss: Enhances the model's focus on extreme wind power points, shifting its focus from optimizing average accuracy to prioritizing the prediction accuracy of important extreme points that affect the determination of ramp events.

[0118] This invention employs a wind power-oriented conformal inference method to construct an adaptive confidence interval. This interval automatically widens in areas of high prediction uncertainty to improve the coverage probability of extreme points in WPREs and effectively reduce the risk of missed detections. After constructing the confidence interval, we further designed a confidence-aware ramp detection algorithm, which uses the upper and lower bounds of the confidence interval to determine whether WPREs will occur in the future, and introduces an adaptive wind power change threshold to reduce false alarms. The algorithm compares the difference between the maximum upper bound and the minimum lower bound in the prediction interval, and then combines this with a dynamic threshold to determine whether a ramp event exists, thereby obtaining more reliable day-ahead ramp prediction results.

[0119] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting day-ahead wind power ramp-up events, characterized in that: Includes the following steps: S1. Day-ahead wind power forecast: Using an extreme value driven model, based on historical wind power and meteorological characteristics, wind power forecasts are generated for Q time points on the next day; S2. Construction of Confidence Intervals for Prediction: Using the conformal inference method guided by wind power, a confidence interval C is constructed for each future time point; S3. Confidence Interval-Based Ramp-up Event Detection: The prediction results of the wind power ramp-up event for the next day are obtained through a confidence-aware detection algorithm.

2. The method for predicting day-ahead wind power ramp-up events according to claim 1, characterized in that: In step S1, the day-ahead wind power forecast includes: S101. Construct an extremum-driven model, including a feature embedding module, an encoder, and a projection layer; The feature embedding module is used to convert historical data into feature representations suitable for deep learning models, including variational mode decomposition units, embedding networks, and embedding units. The encoder contains an L-layer structure, with each layer encoded by a HFFE module. It uses an attention mechanism to extract short-term and long-term dependency structures. The HFFE module refers to the Hybrid Frequency Feature Enhancement Module. The projection layer decodes the encoder output into a daily wind power prediction. S102. Collect historical data of the target wind farm and construct a database of the target wind farm, including wind power data of N wind turbines in the wind farm at the past K time points, meteorological characteristics W composed of G types of weather in the wind farm, and spatial adjacency matrix A of spatial correlation between N wind turbines. S103. For any wind turbine in the wind farm, construct a training dataset: (1) From the wind power data of the wind turbine at the past K time points, select wind power data for a consecutive M days. Let the number of time points in each day be Q. The wind power data of the first M-1 days in the consecutive M days constitute the power sequence X(t) of the current wind turbine, forming the input sample. The actual wind power data of the Q time points in the Mth day are used as labels. The features of the input sample and the corresponding labels constitute a training sample, where M>2 and M*Q <K; (2) Repeat (1) to obtain multiple training samples, which constitute the training sample set of the current wind turbine; S104. Train the extreme value driving model of any wind turbine in the wind farm to obtain the trained extreme value driving model corresponding to the wind turbine. S105. For each wind turbine in the wind farm, repeat steps S103 to S104 to obtain the trained extreme value driving model corresponding to each wind turbine in the wind farm. S106. For each wind turbine in the wind farm, construct a power sequence using the wind power data from the most recent M-1 days, input it into the trained extreme value driving model corresponding to the wind turbine, and obtain the wind power prediction for Q time points on the next day.

3. The method for predicting day-ahead wind power ramp-up events according to claim 2, characterized in that: Step S104 includes: For any wind turbine in a wind farm, based on each sample in the training sample set of that wind turbine, the power sequence X(t) is input into the feature embedding module for processing, and the historical data is converted into a feature representation suitable for deep learning model processing, thus forming the embedded feature. The output of the feature embedding module is encoded by the encoder, and then the encoded result is sent to the projection layer. In the projection layer, the power prediction values ​​for the next Q time points are obtained by decoding through a multilayer perceptron. The loss function is calculated using the power predictions for the next Q time points and the labels of the current samples, and the extreme value-driven model is updated accordingly. Once training has been completed for each sample in the training sample set, or the loss function is less than a set threshold, the resulting extreme value-driven model is used as the extreme value-driven model for the current wind turbine.

4. The method for predicting day-ahead wind power ramp-up events according to claim 3, characterized in that: The process of inputting the power sequence X(t) into the feature embedding module converts historical data into feature representations suitable for deep learning models, forming embedded features, including: Capturing multi-scale time series features: Using variational mode decomposition units, the power sequence X(t) of each wind turbine is decomposed into M intrinsic IMF components, which constitute IMF features; The decomposed IMF features, spatial adjacency matrix A, meteorological features W, and power sequence X(t) are fed into the embedding network and mapped to a unified D-dimensional feature space to obtain the mapping result of the IMF features. Mapping results of spatial adjacency matrix Mapping results of meteorological characteristics , and the mapping results of the power sequence; Spatiotemporal embedding is generated through embedding units. Space-meteorological embedding : ； ； This yields the output of the embedded unit. .

5. The method for predicting day-ahead wind power ramp-up events according to claim 4, characterized in that: The process of encoding the output of the feature embedding module using an encoder includes: The encoder contains an L-layer structure. Each layer encodes through an HFFE block. A feedforward network and normalization processing are added to each HFFE block. The fused features are then passed through another feedforward network and normalization layer to complete the processing of an HFFE block. Each layer's HFFE block aims to capture both long and short-term dependencies simultaneously, that is, to capture short-term changes and long-term advantages through frequency attention and feature attention. The processing procedure for the HFFE block in layer l is as follows: For spatiotemporal embedding By calculating its attention weights on different IMF frequency components through frequency attention, it focuses on short-term, sharp fluctuations associated with sudden changes: ； in, This represents the hidden state of the l-th layer. This represents the frequency attention calculation result of the l-th layer; For space-weather embedding The attention weights of the impact of different meteorological characteristics on power are calculated to provide long-term trend information. ； in, This represents the result of the feature attention calculation for the l-th layer; The two attention outputs are adaptively fused using relevant decision gates. : ； in, , As a weight for relevant decision-makers, The bias of the relevant decision-makers , This represents element-wise multiplication; The hidden state of the encoder's first layer is the mapping result of the power sequence. From the second layer to the (L-1)th layer, each layer obtains... After passing through the feedforward network and the normalization layer, it serves as the hidden state for the next layer, resulting in the Lth layer. The output of the encoder is obtained by passing through a feedforward network and a normalization layer.

6. The method for predicting day-ahead wind power ramp-up events according to claim 3, characterized in that: The extreme value-driven model enhances the prediction accuracy of extreme points through extreme value loss, the loss function of which is defined as: ； Here, the predicted time range, i.e., the Q time points of the next day, are divided into multiple time windows. This represents the set of time points representing extreme values ​​within each time window. This indicates the number of extreme points; the extreme values ​​refer to the maximum and minimum values ​​within a time window, and the time points of the extreme values ​​refer to the times when the extreme values ​​occur within each time window. The weights of the extreme value terms; Q represents the number of time points within the predicted time range. , This represents the actual and predicted wind power output at time point t. The Adam optimizer or stochastic gradient descent implementation used when updating the extreme value driven model.

7. The method for predicting day-ahead wind power ramp-up events according to claim 1, characterized in that: Step S2 includes: S201. Based on the extreme value driving model of each wind turbine, predict the wind power of the wind turbine in the most recent two days and the next day; at each predicted time point, superimpose the predicted power of each wind turbine in the wind farm to obtain the predicted wind power value of the wind farm in the most recent two days and the next day. By superimposing the actual power of each wind turbine in the wind farm at each point in time within the most recent two days in history, the true value of wind power at each point in time within the most recent two days in history is obtained. The prediction period is three days, with Q time points per day, for a total of 3Q time points. The most recent two historical time points are denoted as follows: ~ Record the time of the next day as ~ ; Construct a calibration set, including: Forecast data This indicates that within the prediction interval, the wind farm starts from the first time point. By the 2Qth time point The set of predicted wind power values; Actual data This indicates that within the prediction interval, the wind farm starts from the first time point. By the 2Qth time point The set of actual wind power values; S202. Calculate the absolute value of the prediction residual for each time point in the calibration set prediction interval, as the conformal fraction: ； in, Let represent the conformal fraction at the t-th time point in the prediction interval. , This represents the actual wind power output and the predicted wind power output at time point t. S203. Conformal fractions in the calibration set Apply time decay weights : ； For future time points If i is not greater than Q, use the conformal fraction of the history window. Calculate the weighted 1-α quantile: ； in, This indicates the time point from the i-th time in the prediction interval. up to the (Q+i)th time point The set of conformal fractions; Indicates to The calculated 1-α quantile; α represents the pre-given confidence level; right Select conformal fractions in ascending order until the normalized weights corresponding to the selected conformal fractions are reached. The sum is greater than or equal to 1-α, for the selected conformal fraction, and its weight Multiplication Set to 1; for unselected conformal fractions, multiply by their weights. Set to 0; Used to handle extreme cases where the fractional digit condition cannot be met, when After selecting all conformal fractions, normalize the weights. If the sum is still less than 1-α, then It is 1 if it is true, otherwise it is 0; in, ; ; S203. Construct confidence intervals: future time points Construct upper and lower bounds to obtain the confidence interval C: ； The upper bound is "+", and the lower bound is "-". This represents the wind power data at the 2Q+i time point predicted using the extreme value driven model; Simultaneously, physical feasibility constraints are incorporated: ； That is, the lower world Cannot be lower than 0, upper bound The power output must not exceed the feasible power range of the wind turbine. ; The upper and lower bounds are calculated on a rolling basis as time changes.

8. The method for predicting day-ahead wind power ramp-up events according to claim 1, characterized in that: Step S3 includes: S301. Define dynamic threshold: Adaptive threshold To reduce false alarms, the threshold for each sliding window of two hours is defined as follows: ； in, Based on a fixed threshold; Indicates the width of the confidence interval at time point t; This indicates the average width of the interval within the window. This represents the average interval width over a day, specifically within Q time points. This is the starting point of the sliding window. The length of the sliding window is set to two hours. S302. Using confidence intervals for hill climb detection: Slide one up the time series of the predicted day For each time window of a given length, calculate the maximum value of the upper bound of its built-in confidence interval. minimum value of the lower bound : ； ； in, It is the set of upper and lower bounds of the confidence interval for each time point within the time window; If satisfied If a wind power ramp-up event occurs in the current window, then determine that a wind power ramp-up event has occurred and output the result. Otherwise, output 0 and continue sliding the window.

Images