A wind power prediction method based on reference power curve state residual learning

Abstract

The application discloses a wind power prediction method based on reference power curve state residual learning. The method cleans, repairs and smooths the historical data of the wind turbine to obtain a normal operation sample set. Based on the sample set, the wind speed is divided into bins, the representative wind speed value, the representative power value and the local dispersion are extracted, the reference power curve is constructed, and the physical reference power, the local dispersion and the local sensitivity corresponding to the target time are obtained. According to the actual power and the physical reference power, the original residual is constructed, and the normalized representation is performed. The normalized residual and the related state characteristics are constructed into a reference guide time sequence input sequence, input into a BiMamba network for bidirectional time sequence modeling, and the normalized residual prediction value is output. The wind power prediction result is obtained by inverse normalization and reference power reconstruction. The application can improve the consistency of residual representation in different wind speed intervals, and enhance the accuracy and interpretability of wind power prediction.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation prediction technology, specifically involving a wind power prediction method based on the state residual learning of a reference power curve. Background Technology

[0002] During actual operation, the output power of wind turbines fluctuates significantly with changes in wind speed. Furthermore, the actual power output under the same wind speed conditions can deviate due to meteorological disturbances, changes in turbine operating status, and measurement noise, resulting in strong instability and randomness in wind power sequences. Existing wind power prediction methods often struggle to simultaneously address these issues by accurately representing the fundamental wind speed-power relationship and characterizing complex disturbances and deviations. This often leads to problems such as mixed modeling of fundamental trends and disturbances, and insufficient stability of prediction results.

[0003] Within different wind speed ranges, the physical meaning of power deviations of the same order of magnitude is not consistent. In ranges where the power curve is relatively flat and the samples are relatively concentrated, small deviations may already indicate significant deviations, while in ranges where the power curve is relatively steep or the samples themselves are relatively dispersed, the same deviation may still be within the normal fluctuation range. Furthermore, some existing deep time series prediction methods lack a clear intermediate representation layer between the front-end reference pattern construction and the back-end neural network input. They fail to unify the reference power, local dispersion, local sensitivity, and residual objects into a unified input representation that can be learned by the sequence model, resulting in unclear model input logic and insufficient physical interpretability.

[0004] In recent years, Mamba-like networks based on state-space models have attracted attention in the field of long sequence modeling. These methods represent sequence information through selective state-space modeling mechanisms, demonstrating good modeling capabilities and computational efficiency when processing long-term sequences. Addressing the characteristic of the original Mamba network, which primarily propagates information along a single temporal direction, previous research has proposed BiMamba-like methods. These methods jointly model the input sequence in both forward and backward directions to enhance the representation of temporal correlations. While these methods offer a new technical approach for time-series dependency modeling in wind power forecasting, existing methods largely focus on extracting general time-series patterns, and the utilization of the physical constraint information inherent in the reference power curve remains insufficient. Furthermore, a unified modeling solution is lacking, encompassing curve state variables, residual normalization representation, and power reconstruction. Summary of the Invention

[0005] This invention provides a wind power prediction method based on state residual learning of reference power curves, which addresses the problems of insufficient connection between reference power and residual modeling, weak residual representation ability in different wind speed ranges, and unclear model input organization in existing wind power prediction methods, thereby improving the accuracy, stability, and rationality of wind power prediction.

[0006] A wind power prediction method based on state residual learning of a reference power curve includes the following steps:

[0007] Step S1: Obtain the target wind power dataset, check the validity of the original wind turbine operation data, use the CATCH anomaly detection algorithm and SAITS unified compensation model to repair outliers and missing values, and use the moving average method to smooth and denoise the data to obtain the normal operation sample set.

[0008] Step S2: Divide the normal operation sample set into bins according to preset wind speed intervals, extract the representative wind speed value, representative power value and local dispersion of each wind speed interval, and construct the wind speed-power mapping relationship based on the representative points through piecewise linear interpolation; use the wind speed value corresponding to the target time to determine the physical reference power and local dispersion of the target time, determine the local sensitivity according to the slope of the interpolation interval, and impose boundary constraints on the physical reference power;

[0009] Step S3: Construct the original residuals based on the actual power of the historical samples and the physical reference power at the corresponding time; determine the normalization scale based on the local dispersion and local sensitivity at the corresponding time, and use the normalization scale to normalize the original residuals to obtain the normalized residual sequence.

[0010] Step S4: Align the normalized residual sequence with the physical reference power, local dispersion, local sensitivity, and wind speed and wind speed change at the corresponding time step by time step, and then perform feature splicing. Use a sliding time window to construct historical time series input samples, and obtain the reference guidance time series input sequence through unified dimension mapping.

[0011] Step S5: Input the reference-guided time series input sequence into the forward Mamba branch and the backward Mamba branch respectively, extract the bidirectional time series features of the target time, and obtain the bidirectional time series representation through cascade fusion; use the bidirectional time series representation and the physical reference power, normalized scale and wind speed change corresponding to the target time to output the normalized residual prediction value; then perform inverse normalization according to the normalized scale of the corresponding time to obtain the prediction residual, and reconstruct it with the corresponding physical reference power to obtain the wind power prediction result of the target time.

[0012] Compared with existing technologies, the wind power prediction method provided by this invention has the following advantages:

[0013] (1) The physical reference power curve is constructed by wind speed sub-box and piecewise linear interpolation, which can extract the basic wind speed-power law that is more consistent with the actual operating conditions of the unit from historical normal operation samples.

[0014] (2) Local dispersion and local sensitivity are normalized to the original residuals so that the residuals under different wind speed ranges can be represented under a unified interpretation scale, thereby reducing the adverse effects of inconsistent residual fluctuation levels in different wind speed ranges on subsequent time series model training.

[0015] (3) The normalized residual, reference power, local dispersion, local sensitivity and wind speed characteristics are jointly constructed as a reference-guided time series input sequence. At the input level, the physical constraint information of the reference power curve and the residual change information can be retained at the same time, which is convenient for subsequent time series models to make predictions in combination with physical reference laws and historical residual states.

[0016] (4) The BiMamba network is used to extract bidirectional time series features from the reference-guided time series input sequence, which can simultaneously characterize the forward and backward correlation information of the target time. On this basis, the residual prediction result is restored to the target power prediction result through inverse normalization and reference power reconstruction, which is conducive to enhancing the physical correspondence between the prediction process and the output result and improving the applicability of the method in actual wind power prediction scenarios. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0018] Figure 1 This is a flowchart illustrating the wind power prediction method of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0021] The present invention establishes a wind power prediction model based on a reference power curve. Historical normal operation samples are used to construct the reference power curve, and local dispersion and local sensitivity are extracted from the reference power curve to characterize the curve state of different wind speed intervals. The original residual is normalized and characterized under the constraint of the curve state quantity, the normalized residual and the reference guidance feature are unified into a time-series input sequence, and the bidirectional Mamba network learns the bidirectional time-series correlation of the residual accordingly. The normalized residual prediction value output by the model is subjected to inverse normalization processing, and then power reconstruction is completed together with the reference power, so as to obtain the wind power prediction value at the target moment.

[0022] The present invention will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only used to illustrate the present invention, not to limit the protection scope of the present invention.

[0023] Please refer to Figure 1 , the wind power prediction method provided by the embodiment of the present invention includes step S1 to step S5:

[0024] In step S1, validity check is performed on the original wind turbine operation data; the CATCH anomaly detection algorithm is used to identify abnormal values, the abnormal values and missing values are uniformly constructed into positions to be repaired, and the SAITS unified value filling model is used for completion and repair. Further, a moving average method is used for smoothing and denoising the repaired time-series data to obtain a normal operation sample set, and the moving average method is expressed as:

[0025]

[0026] wherein, represents the smoothed data value at time , represents the original data value at time , when , represents the original data value at time , represents the length of the smoothing window, represents the summation index in the moving average window.

[0027] In step S2, wind speed binning is performed on the normal operation sample set, and the representative wind speed value, representative power value and local dispersion index of each wind speed interval are extracted, and then the physical reference power, local dispersion and local sensitivity at the target moment are obtained through piecewise linear interpolation. The method can not only establish a reference power conforming to the historical operation law of the unit, but also further characterize the local state of the reference power curve under different wind speed intervals, providing a basis for the subsequent scaled representation of residuals.

[0028] Step S2 specifically includes:

[0029] Step S201: Divide the normal operation sample set into boxes according to the preset wind speed and width. Separate boxes, the first Each wind speed range is represented as:

[0030]

[0031] in, Indicates the lower limit of wind speed. This indicates the total number of wind speed ranges.

[0032] Step S202: Extract representative wind speed values ​​within each wind speed range using the median method. and represent power values :

[0033]

[0034]

[0035] in, Indicates the first Wind speed values ​​for each sample Indicates the first The actual power value corresponding to each sample This indicates median operations.

[0036] Since the median reflects the central position of a sample sequence, even if a small number of outliers deviate significantly from the median, the median result is usually not significantly changed. Therefore, when real-world data contains local disturbances, instantaneous fluctuations, and individual outliers, using the median as a representative statistic can reduce the interference of outliers on the statistical results and improve the stability and reliability of the representative value.

[0037] The power curve discrete point set is composed of representative wind speed values ​​and representative power values ​​for each wind speed range. :

[0038]

[0039] Step S203: Calculate the local dispersion index of the power samples within each wind speed range.

[0040]

[0041] in, and Let represent the upper quartile and lower quartile of the power sample within the m-th wind speed interval, respectively.

[0042] The upper and lower quartiles are used to characterize the statistical levels of the higher and lower positions in the power sample distribution within the wind speed range, respectively. The difference between the two reflects the distribution span of the middle 50% of the power samples within the wind speed range. Therefore, It can be used to describe the local fluctuation of power samples within the current wind speed range. Compared with directly using the difference between the maximum and minimum values, the interquartile range is less affected by extreme abnormal samples and has better robustness.

[0043] This constitutes the discrete point set. :

[0044]

[0045] Among them, local dispersion The larger the value, the more dispersed the power samples are within that wind speed range, indicating higher local dispersion. The smaller the value, the more concentrated the power samples are within that wind speed range.

[0046] Step S204, set the wind speed value corresponding to the target time as follows: When satisfied At that time, piecewise linear interpolation is used to determine the physical reference power at the target time. and local dispersion :

[0047]

[0048]

[0049] in, Indicates the first Representative power values ​​for each wind speed range. Indicates the first The representative wind speed value for each wind speed range.

[0050] Local sensitivity is determined based on the interpolation interval in which the target time lies. This is used to characterize the local slope of the current reference power curve within this wind speed range:

[0051]

[0052] Boundary constraints are applied to the physical reference power:

[0053]

[0054] in, This indicates the rated power of the wind turbine.

[0055] In step S3, the original residuals are constructed based on the actual power and the physical reference power. The normalization scale is then determined by combining local dispersion and local sensitivity, and the original residuals are normalized to obtain a normalized residual sequence. In this way, residuals with inconsistent meanings in different wind speed ranges can be represented on a unified scale, thereby reducing the impact of local fluctuations on model training and enabling subsequent models to learn the residual variation patterns more effectively.

[0056] Step S3 is as follows:

[0057] Step S301: Construct the original residuals based on the actual power of the historical samples and the physical reference power at the corresponding time. :

[0058]

[0059] in, Indicates time The actual power Essentially, it is used to represent the deviation of actual power from reference power.

[0060] Step S302: Determine the normalization scale based on the local dispersion and local sensitivity at the corresponding time. :

[0061]

[0062] in, This represents the natural dispersion of the power sample itself within the current wind speed range. This indicates the typical power variation scale corresponding to the local slope of the current reference power curve within the current bin width range.

[0063] therefore It can be used to characterize a reasonable fluctuation scale of the current residual.

[0064] Step S303: Normalize the original residuals using a normalized scale to obtain the curve state normalized residuals. :

[0065]

[0066] Among them, the normalization scale When the value is greater than 0, the original residual is scaled.

[0067] Normalized scale When the denominator is zero, the original residual is retained directly to avoid numerical anomalies caused by a zero denominator. This processing allows the residuals across different wind speed ranges to be characterized on a more uniform interpretive scale.

[0068] In step S4, the normalized residuals, physical reference power, local dispersion, local sensitivity, and wind speed and wind speed change at corresponding times are aligned and concatenated according to time steps. A reference-guided time-series input sequence is constructed using a sliding time window and a unified dimension mapping. This input sequence uniformly includes residual characterization information, reference power information, and curve state information, facilitating the BiMamba model to model the temporal correlations in the residual evolution process and to incorporate the corresponding reference curve state information.

[0069] Step S4 is as follows:

[0070] Step S401, for each time step Constructing reference-guided feature vectors :

[0071]

[0072] in, This represents the change in wind speed. The main information related to power deviation at a single time step is organized into a unified feature vector, thus forming the basis for constructing subsequent time-series inputs.

[0073] Step S402, based on the length of Constructing historical time series input samples using a sliding time window:

[0074]

[0075] Specifically, the reference guidance feature vector of a single time step is expanded into a sequence sample containing historical context information, enabling the BiMamba model to learn the dynamic evolution of the normalized residual and its reference curve state information over continuous time, rather than making a static mapping based on a single moment.

[0076] Step S403: Perform a unified dimensional mapping on the features at each time step to obtain the temporal coding vector:

[0077]

[0078] in, This represents the sequence position index within the sliding time window. For the mapping weight matrix, This represents the reference guiding feature vector corresponding to a certain time step in the historical time series input sample. For bias terms, This is the activation function.

[0079] By mapping input quantities with different dimensions and physical meanings in the original feature vectors to a unified feature space, the BiMamba model can perform temporal dependency modeling within a consistent representation space.

[0080] The reference bootstrap timing input sequence is composed of the timing coding vectors at each time step:

[0081]

[0082] In step S5, the reference-guided time-series input sequence is input into the BiMamba model, the forward and reverse time-series dependency features are extracted, the normalized residual prediction value at the target time is output, and the final wind power prediction result is obtained by inverse normalization and reconstruction with the physical reference power.

[0083] Step S5 is as follows:

[0084] Step S501, input the reference boot timing sequence Input a BiMamba model. The forward Mamba unit scans the reference-guided temporal input sequence along the forward time axis to obtain the forward temporal output features. The backward Mamba unit scans the reference guiding timing input sequence in reverse time to obtain the backward timing output features. :

[0085]

[0086]

[0087] in, This represents the temporal feature extraction function for the forward Mamba branch. This represents the temporal feature extraction function for the backward Mamba branch.

[0088] The forward Mamba branch focuses more on the evolution trend of the normalized residuals and reference curve state variables from historical moments to the current moment, while the backward Mamba branch focuses more on supplementing the overall constraint information of the local segment where the target moment is located from the reverse perspective. By simultaneously introducing the temporal features of the above two directions, the model's ability to represent complex wind power fluctuation patterns can be enhanced.

[0089] Step S502: Concatenate the forward and backward time series output features and map them into a bidirectional time series representation of the target time:

[0090]

[0091] in, Indicates the target time Bidirectional timing representation, This indicates a feature cascade operation. Represents the fusion mapping matrix, This indicates the bias term.

[0092] By mapping forward temporal dependency features and backward constraint features to the same feature space, a representation result containing bidirectional temporal information is obtained. Compared with unidirectional temporal modeling alone, bidirectional temporal representation can more fully characterize the dynamic correlation between normalized residuals and curve state variables in the reference-guided temporal input sequence.

[0093] Step S503, set the target time The corresponding physical reference power, normalized scale, and wind speed variation are used to construct a curved state vector. :

[0094]

[0095] Based on bidirectional time series representation and curved state vectors, the normalized residual prediction value at the target time is output. :

[0096]

[0097] in, Represents the prediction mapping matrix, This indicates the bias term.

[0098] The normalized residual prediction not only retains BiMamba's extraction results of time dependence, but also further incorporates the local state information of the reference power curve corresponding to the target time, so that the prediction results can simultaneously reflect the temporal evolution law and physical state constraints.

[0099] Step S504: Based on the normalized scale constructed in step S3, perform inverse normalization on the predicted normalized residual values ​​to obtain the predicted residuals:

[0100]

[0101] Among them, the normalization scale At that time, through Complete inverse normalization recovery; when This indicates that there is no effective scaling effect at the current moment, and the normalized residual prediction value is directly applied. This is used as the output of the prediction residual. This processing method is consistent with the aforementioned normalization definition and can avoid numerical anomalies caused by invalid scaling.

[0102] The predicted residuals are reconstructed with the corresponding physical reference power to obtain the final wind power prediction result:

[0103]

[0104] in, This represents the predicted wind power output at the target time.

[0105] And perform physical boundary correction on the prediction results:

[0106]

[0107] in, This indicates the rated power of the wind turbine.

[0108] To verify the effectiveness of the method of this invention, historical data from a wind farm in Xinjiang, China, were selected for comparative and ablation experiments. The samples were divided into training, validation, and test sets according to time sequence, and the goodness-of-fit R-squared was used. 2 The mean absolute error (MAE) and root mean square error (RMSE) were used as evaluation indicators, and the comparison results are shown in Table 1.

[0109] The experiment included the following four sets of models:

[0110] Model (A) of this invention: adopts a reference power curve state residual learning framework, constructs a reference power curve using historical normal operation samples, extracts local dispersion and local sensitivity, normalizes the original residual, and constructs a reference-guided time series input sequence using the normalized residual, physical reference power, local dispersion, local sensitivity, wind speed and wind speed variation, and inputs it into a BiMamba network for bidirectional time series modeling. Finally, the wind power prediction result is obtained through inverse normalization and reference power reconstruction.

[0111] The basic BiMamba model (B) directly uses the BiMamba network to predict the input sequence without introducing a reference power curve, local dispersion, local sensitivity, or residual normalization mechanism. It is used to verify the effect of the reference power curve state residual learning idea proposed in this invention on improving prediction performance.

[0112] One-way Mamba Alternative Model (C): Retains the reference power curve construction, curve state quantity extraction, residual normalization and reference time series input construction methods in this invention, and only uses one-way Mamba to replace BiMamba for modeling, in order to verify the impact of bidirectional time series dependency modeling on prediction performance.

[0113] BiLSTM Alternative Model (D): Retains the reference power curve construction, curve state quantity extraction, residual normalization and reference time series input construction methods in this invention, and only replaces BiMamba with BiLSTM for modeling, to verify the compatibility between the state residual learning framework proposed in this invention and the BiMamba network.

[0114] Model <![CDATA[R 2 ]]> MAE RMSE A 0.995 0.279 0.351 B 0.984 0.468 0.596 C 0.990 0.336 0.421 D 0.988 0.358 0.446

[0115] As shown in Table 1, Model A outperforms the other comparative models in all evaluation indicators, with R... 2The R-value reached 0.995, the highest among the four models, while the MAE and RMSE were 0.279 and 0.351, respectively, both the lowest, indicating that the method of this invention has the best prediction accuracy and fitting effect; compared with model B, the R-value of the model of this invention is higher. 2 The MAE and RMSE decreased from 0.468 and 0.596, respectively, to 0.279 and 0.351, respectively, indicating that the reference power curve state residual learning mechanism can effectively improve the model's prediction performance. Compared with Model C, the model of this invention achieves even better results, demonstrating that bidirectional temporal dependency modeling helps enhance the representation of residual change characteristics at the target time. Compared with Model D, the model of this invention also maintains a significant advantage, indicating that under the same state residual learning framework, the BiMamba model and the method of this invention have better adaptability. In summary, the method of this invention shows better performance in terms of prediction accuracy, temporal modeling capability, and overall applicability.

[0116] In summary, the wind power prediction method provided by this invention has the following beneficial effects:

[0117] (1) The physical reference power curve is constructed by wind speed sub-box and piecewise linear interpolation, which can extract the basic wind speed-power law that is more consistent with the actual operating conditions of the unit from historical normal operation samples.

[0118] (2) Local dispersion and local sensitivity are normalized to the original residuals so that the residuals under different wind speed ranges can be represented under a unified interpretation scale, thereby reducing the adverse effects of inconsistent residual fluctuation levels in different wind speed ranges on subsequent time series model training.

[0119] (3) The normalized residual, reference power, local dispersion, local sensitivity and wind speed characteristics are jointly constructed as a reference-guided time series input sequence. At the input level, the physical constraint information of the reference power curve and the residual change information can be retained at the same time, which is convenient for subsequent time series models to make predictions in combination with physical reference laws and historical residual states.

[0120] (4) The BiMamba network is used to extract bidirectional time series features from the reference-guided time series input sequence, which can simultaneously characterize the forward and backward correlation information of the target time. On this basis, the residual prediction result is restored to the target power prediction result through inverse normalization and reference power reconstruction, which is conducive to enhancing the physical correspondence between the prediction process and the output result and improving the applicability of the method in actual wind power prediction scenarios.

[0121] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention or equivalent substitutions can be made to some of the technical features therein, without departing from the spirit and scope of the technical solutions of the present invention.

[0122] It should be understood that the sequence number, name, and order of the steps in the above embodiments are only for the convenience of describing the implementation process of the present invention and do not constitute a limitation on the implementation order of the present invention. In specific implementation, the execution order of each step can be adjusted accordingly based on the actual application scenario, data processing flow, and functional logic.

[0123] For those skilled in the art, various modifications, improvements, and substitutions can be made without departing from the concept of the present invention, and all such modifications, improvements, and substitutions should fall within the protection scope defined by the claims of the present invention.

Claims

1. A wind power prediction method based on state residual learning of a reference power curve, characterized in that, Includes the following steps: Step S1: Obtain the target wind power dataset, check the validity of the original wind turbine operation data, use the CATCH anomaly detection algorithm and SAITS unified compensation model to repair outliers and missing values, and use the moving average method to smooth and denoise the data to obtain the normal operation sample set. Step S2: Divide the normal operation sample set into bins according to preset wind speed intervals, extract the representative wind speed value, representative power value and local dispersion of each wind speed interval, and construct the wind speed-power mapping relationship based on the representative points through piecewise linear interpolation; use the wind speed value corresponding to the target time to determine the physical reference power and local dispersion of the target time, determine the local sensitivity according to the slope of the interpolation interval, and impose boundary constraints on the physical reference power; Step S3: Construct the original residuals based on the actual power of the historical samples and the physical reference power at the corresponding time; determine the normalization scale based on the local dispersion and local sensitivity at the corresponding time, and use the normalization scale to normalize the original residuals to obtain the normalized residual sequence. Step S4: Align the normalized residual sequence with the physical reference power, local dispersion, local sensitivity, and wind speed and wind speed change at the corresponding time step by time step, and then perform feature splicing. Use a sliding time window to construct historical time series input samples, and obtain the reference guidance time series input sequence through unified dimension mapping. Step S5: Input the reference guiding time series input sequence into the forward Mamba branch and the backward Mamba branch respectively, extract the bidirectional time series features of the target time, and obtain the bidirectional time series representation through cascade fusion; use the bidirectional time series representation and the physical reference power, normalized scale and wind speed change corresponding to the target time to output the normalized residual prediction value. Then, inverse normalization is performed based on the normalized scale at the corresponding time to obtain the prediction residual, which is then reconstructed with the corresponding physical reference power to obtain the wind power prediction result at the target time.

2. The wind power prediction method according to claim 1, characterized in that, Step S1 specifically includes: The original wind turbine operating data undergoes validity checks; outliers are identified using the CATCH anomaly detection algorithm, and these outliers and missing values ​​are uniformly constructed as locations to be repaired. The SAITS unified imputation model is then used for imputation and repair. Furthermore, the repaired time-series data is smoothed and denoised using a moving average method to obtain a normal operating sample set. The moving average method is expressed as follows: wherein, represents the smoothed data value at time , represents the original data value at time , when , represents the original data value at time , represents the length of the smoothing window, represents the summation index within the sliding average window.

3. The wind power prediction method according to claim 1, characterized in that, Step S2 specifically includes: Step S201: Divide the normal operation sample set into boxes according to the preset wind speed and width. Separate boxes, the first wind speed range Represented as: in, Indicates the lower limit of wind speed. Indicates the total number of wind speed ranges; Step S202: Extract representative wind speed values ​​within each wind speed range using the median method. and represent power values : in, Indicates the first Wind speed values ​​for each sample Indicates the first The actual power value corresponding to each sample This indicates median operations. The power curve discrete point set is composed of representative wind speed values ​​and representative power values ​​for each wind speed range. : Step S203: Statistically analyze the local dispersion index of power samples within each wind speed range. : in, and Let represent the upper quartile and lower quartile of the power sample within the m-th wind speed interval, respectively. This constitutes the discrete point set. : Step S204, set the wind speed value corresponding to the target time as follows: ,when At that time, piecewise linear interpolation is used to determine the physical reference power at the target time. and local dispersion : in, Indicates the first Representative power values ​​for each wind speed range. Indicates the first The representative wind speed value for each wind speed range. Local sensitivity is determined based on the interpolation interval in which the target time lies. : Boundary constraints are applied to the physical reference power: in, This indicates the rated power of the wind turbine.

4. The wind power prediction method according to claim 1, characterized in that, Step S3 specifically includes: Step S301: Construct the original residuals based on the actual power of the historical samples and the physical reference power at the corresponding time. : in, Indicates time The actual power. Step S302: Determine the normalization scale based on the local dispersion and local sensitivity at the corresponding time. : in, This represents the natural dispersion of the power sample itself within the current wind speed range. This indicates the typical power variation scale corresponding to the local slope of the current reference power curve within the current bin width range. Step S303: Normalize the original residuals using a normalized scale to obtain the curve state normalized residuals:

5. The wind power prediction method according to claim 1, characterized in that, Step S4 specifically includes: Step S401, for each time step Constructing reference-guided feature vectors : in, This indicates the change in wind speed. Step S402, based on the length of Constructing historical time series input samples using a sliding time window: Step S403: Perform a unified dimensional mapping on the features at each time step to obtain the temporal coding vector: in, This represents the sequence position index within the sliding time window. For the mapping weight matrix, This represents the reference guiding feature vector corresponding to a certain time step in the historical time series input sample. For bias terms, This is the activation function. The reference bootstrap timing input sequence is composed of the timing coding vectors at each time step:

6. The wind power prediction method according to claim 1, characterized in that, Step S5 specifically includes: Step S501, input the reference boot timing sequence By inputting the forward Mamba branch and the backward Mamba branch respectively, the forward time series output features at the target time are obtained. and backward timing output features : in, This represents the temporal feature extraction function for the forward Mamba branch. This represents the temporal feature extraction function for the backward Mamba branch. Step S502: Concatenate the forward and backward time series output features and map them into a bidirectional time series representation of the target time: in, Indicates the target time Bidirectional timing representation, Indicates feature cascading, Represents the fusion mapping matrix, This indicates the bias term. Step S503: Construct a curved state vector from the physical reference power, normalized scale, and wind speed change corresponding to the target time. : Based on bidirectional time series representation and curved state vectors, the normalized residual prediction value at the target time is output. : in, Represents the prediction mapping matrix, This indicates the bias term. Step S504: Based on the normalized scale constructed in step S3, perform inverse normalization on the predicted normalized residual values ​​to obtain the predicted residuals at the target time. : The predicted residuals are reconstructed with the corresponding physical reference power to obtain the final wind power prediction result: Physical boundary correction is applied to the prediction results: in, This indicates the rated power of the wind turbine.

Images