A wind power prediction method

Abstract

The application provides a wind power prediction method. The method comprises the following steps: obtaining samples and dividing the samples according to seasons; processing the samples; constructing an input feature set; constructing a clustering feature matrix for each seasonal sample and determining a seasonal center, and then removing abnormal samples according to the distance threshold of each seasonal sample to the seasonal center; performing SGMD decomposition to obtain a plurality of modal components; predicting each modal component by using an optimized BiLSTM to obtain a prediction result of each modal component; reconstructing the prediction results of the modal components to obtain an SGMD-WOA-BiLSTM main prediction result; constructing a conservative weight-bias correction model; constructing an LSBoost residual compensation model to realize secondary compensation on the prediction result after conservative weight-bias correction; and performing physical boundary constraint, peak-valley constraint and error gate processing on the compensated prediction result to output a final wind power prediction value. The application greatly improves the prediction accuracy and stability of wind power.

Description

Technical Field

[0001] This invention relates to the technical field of new energy power generation prediction, and more specifically, to a method for predicting wind power. Background Technology

[0002] Wind power output is influenced by multiple factors, including wind speed, wind direction, temperature, air pressure, rainfall, turbine operating status, and terrain, exhibiting significant randomness, intermittency, seasonality, and non-stationarity. High-precision wind power forecasting is crucial for power system dispatching, reserve capacity allocation, wind power integration, and grid connection security.

[0003] Existing wind power forecasting methods are mainly based on models including physical models, statistical models, machine learning models, and deep learning models. Among them, LSTM and BiLSTM can learn the long-term dependencies of time series, but they still have the following problems in actual wind power forecasting: First, when using the original power series directly for modeling, strong non-stationary fluctuations and local abrupt changes in the series increase the difficulty of model learning; second, network hyperparameters rely on human experience, and unreasonable parameter combinations can easily lead to underfitting or overfitting; third, the output of a single neural network is prone to systematically overestimating or underestimating at specific quarters and specific hourly points; fourth, conventional residual compensation usually directly learns the master forecast residual without explicitly correcting the regular amplitude deviation and center bias of the master forecast output, causing the residual model to bear too much correction task.

[0004] Therefore, based on the technical deficiencies in actual wind power prediction, it is necessary to propose a wind power prediction method that can simultaneously perform sample selection, mode decomposition, parameter optimization, master prediction, weight-bias correction, and residual compensation. Furthermore, it is necessary to avoid over-correction through conservative constraints and error gating mechanisms to greatly improve prediction accuracy and stability. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0006] Therefore, the purpose of this invention is to propose a method for predicting wind power.

[0007] To achieve the above objectives, the present invention provides a method for predicting wind power. The wind power prediction method includes: Step S1: Obtaining samples and dividing them into Q1, Q2, Q3, and Q4 according to quarters; wherein the samples include: historical wind power data of wind farms, meteorological data, and time information; Step S2: Performing time standardization, missing value processing, outlier processing, normalization processing, and effective sample screening on the samples; Step S3: Constructing an input feature set; wherein the input feature set includes: temperature fuzzy mapping features, time period features, lag power features, difference features, moving mean features, moving standard deviation features, rolling extreme value features, and meteorological enhancement features; Step S4: Constructing a clustering feature matrix for the samples of each quarter, and using a single-center clustering screening method to determine the quarter center, and then removing abnormal samples far from the center based on the distance threshold from the samples of each quarter to the quarter center; Step S5: Performing SGMD decomposition on the power increment sequence to obtain multiple modal components; Step S6: Optimizing the BiLSTM network using WOA, and utilizing the optimized BiLSTM... Step S7: Reconstruct the prediction results of each modal component to obtain the SGMD-WOA-BiLSTM master prediction result; Step S8: Based on the master prediction residuals of different quarters and different time periods in the training set, learn the corresponding weight correction coefficients and bias correction terms, and construct a conservative weight-bias correction model to achieve parameterized correction of the master prediction result; Step S9: Using the residual between the parameterized correction prediction result and the actual power as the learning object, construct an LSBoost residual compensation model to achieve secondary compensation of the prediction result after conservative weight-bias correction; Step S10: Perform physical boundary constraints, peak-valley constraints and error gating on the compensated prediction result, and output the final wind power prediction value.

[0008] Preferably, in step S3, the temperature fuzzy mapping feature is represented by three membership functions: low temperature, medium temperature, and high temperature, and the calculation method is as follows:

[0009] ;

[0010] In the formula, express Temperature value at time, Indicates the membership degree at low temperature. Indicates the membership degree at medium temperature. Indicates the membership degree at high temperature. express Temperature fuzzy mapping characteristics at any given time;

[0011] The time period characteristics include: hourly period characteristics and weekly period characteristics, calculated as follows:

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] In the formula, express The hour corresponding to the time. express The week number corresponding to the time. Represents hourly sine code. Represents hourly cosine encoding. Represents the sine code for the day of the week. Represents the cosine encoding of the weekday;

[0017] The hysteresis power characteristics include , , , and At least one of the following; the difference feature includes first-order difference. and second-order difference At least one of the following: the moving mean feature, the moving standard deviation feature, and the rolling extreme value feature are used to characterize the average level, fluctuation intensity, and local peak and valley state of the power sequence within a preset time window, respectively; the meteorological enhancement feature includes at least one of wind speed, wind direction, rainfall, maximum wind speed, air pressure, and solar radiation intensity.

[0018] Preferably, in step S4, the single-center clustering screening method is performed quarterly, for the first quarter... A cluster feature matrix is ​​constructed from quarterly samples. and calculate the quarterly center ;

[0019] The quarterly center The calculation method is as follows:

[0020] ;

[0021] For the first For each sample, calculate its Euclidean distance to the quarter center:

[0022] ;

[0023] Based on the distance distribution, the quarterly distance threshold is determined as follows:

[0024] ;

[0025] In the formula, Indicates the first Clustering feature matrix for each quarter, Indicates the first Single cluster center for each quarter Indicates the first The number of samples participating in clustering within a quarter. Indicates the first Euclidean distance from each sample to the quarter center Indicates the first Distance threshold for each quarter, Indicates the quantile threshold;

[0026] when When, retain the sample; when If the sample is found to be an anomaly far from the quarterly center, it will be removed.

[0027] Preferably, .

[0028] Preferably, in step S5, the decomposition object of SGMD is a power increment sequence, which is represented as:

[0029] ;

[0030] In the formula, express Real-time wind power output express Real-time wind power output express Power increment at any moment;

[0031] The power increment sequence was decomposed using SGMD to obtain multiple modal components:

[0032] ;

[0033] In the formula, Indicates the first One modal component, Indicates the number of modal components. This indicates the decomposition of residual terms.

[0034] In this technical solution, by decomposing the power increment sequence rather than the original power sequence, the interference of the long-term trend term on the prediction of short-term fluctuations is reduced, and the model's ability to respond to local power changes is enhanced.

[0035] Preferably, in step S6, WOA is used to optimize the parameters of the BiLSTM network, and the parameter vector to be optimized is represented as:

[0036] ;

[0037] In the formula, Indicates the number of hidden layer units in the BiLSTM. This represents the initial learning rate. Represents the L2 regularization coefficient. Indicates the dropout ratio;

[0038] The objective function of WOA uses a weighted combination of the root mean square error and the mean absolute error of the validation set, expressed as:

[0039] ;

[0040] In the formula:

[0041] ;

[0042] ;

[0043] In the formula, ;

[0044] Preferably, , ;

[0045] WOA minimizes To obtain the optimal parameters:

[0046] ;

[0047] In the formula, This represents the optimal parameter combination for BiLSTM obtained through WOA optimization;

[0048] For each SGMD modal component, the optimized BiLSTM is used for prediction, resulting in:

[0049] ;

[0050] In the formula, Indicates the first Each modal component in Predicted value at time, express Input feature sequence at any time, This represents the BiLSTM parameters optimized by WOA.

[0051] Preferably, in step S7, the prediction results of each modal component are reconstructed into predicted power increment values:

[0052] ;

[0053] Further, the master predicted power is obtained:

[0054] ;

[0055] Alternatively, the main prediction power can be obtained through non-negative weighted fusion:

[0056] ;

[0057] In the formula, This represents the BiLSTM modal reconstruction prediction value. Indicates the hysteresis power characteristic, Indicates the characteristics of the moving average. Indicates the maximum scrolling value characteristic. Indicates the rolling minimum value characteristic; , , , and As weight;

[0058] Weight , , , and satisfy:

[0059] ;

[0060] .

[0061] Preferably, in step S8, the conservative weight-bias correction model is constructed separately for each quarter and time period, for the first... The quarter, the The master forecast results for each time period are corrected and expressed as follows:

[0062] ;

[0063] In the formula, Indicates the first The quarter, the The time period, the first The principal prediction power for each sample This represents the predicted power after weight-bias correction. Indicates the first The quarter, the Weighting correction coefficients for each time period Indicates the first The quarter, the Bias correction term for each time period;

[0064] The training residual is represented as:

[0065] ;

[0066] In the formula, Indicates the first The quarter, the The time period, the first The actual wind power output of a sample This represents the master prediction residual for the corresponding sample;

[0067] The weight correction coefficients and bias correction terms are obtained by fitting the training samples for that quarter and that period:

[0068] ;

[0069] The final weight correction coefficients and bias correction terms are obtained through the amplitude limiting function:

[0070] ;

[0071] ;

[0072] In the formula, Indicates will Limited to Within the range, and These represent the lower and upper limits of the weight correction coefficient, respectively. and These represent the lower and upper limits of the bias correction term, respectively.

[0073] Preferably, the conservative weight-bias correction model further introduces a conservative fusion coefficient. This merges the correction result with the original master prediction result, and is expressed as:

[0074] ;

[0075] Right now:

[0076] ;

[0077] In the formula, This represents the predicted power after conservative weighting and bias correction. Denotes the conservative fusion coefficient, and .

[0078] Preferably, the weight correction coefficient satisfies:

[0079] ;

[0080] The conservative fusion coefficient satisfies:

[0081] ;

[0082] By using the conservative fusion coefficients and limiting function, the weight-bias correction model only makes small, controlled corrections to the main prediction results, avoiding over-correction that could cause the prediction results to deviate from the original trend. Preferably, in step S9, the LSBoost residual compensation model constructs residuals based on the conservative weight-bias correction prediction results, and these residuals are expressed as follows:

[0083] ;

[0084] Constructing residual compensation input features:

[0085] ;

[0086] In the formula, This represents the prediction value of a single BiLSTM base model. Indicates the hysteresis power characteristic, Indicates the characteristics of the moving average. Indicates differential power characteristics. Indicates the maximum scrolling value characteristic. Indicates the rolling minimum value characteristic. Represents a set of meteorological enhancement features;

[0087] Learning residual mappings using LSBoost:

[0088] ;

[0089] The final compensated prediction value is:

[0090] ;

[0091] In the formula, Indicates the first The residual scaling factor for each quarter This represents the residual compensation value obtained from LSBoost prediction.

[0092] Preferably, in step S10, the peak-valley constraint includes: valley constraint, peak constraint, and super-peak constraint;

[0093] The physical boundary constraints are expressed as follows:

[0094] ;

[0095] In the formula, This indicates the reasonable upper limit of wind power output;

[0096] The trough constraint, peak constraint, and superpeak constraint are respectively represented as follows:

[0097] ;

[0098] ;

[0099] ;

[0100] In the formula, Indicates the trough threshold. Indicates the peak threshold. Indicates the peak value threshold. Indicates the upper limit of the trough. Indicates the upper bound of the peak value. Indicates the upper bound of the superpeak value;

[0101] The error gating process is expressed as follows:

[0102] ;

[0103] In the formula, This represents the root mean square error of the optimization model. This represents the root mean square error of a single BiLSTM basis model. Indicates the error gating threshold. This represents the fallback fusion coefficient.

[0104] In this technical solution, the physical boundary constraints, peak-valley constraints, and error gating mechanisms are used to limit unreasonable overshoot in the prediction results. When the error of the optimized model is greater than the error of the base model, back-off fusion is performed to improve the stability and engineering usability of the wind power prediction results.

[0105] Compared with existing technologies, the wind power prediction method provided by this invention has the following advantages:

[0106] (1) By using quarterly single-center clustering threshold screening, the main distribution samples are retained within each quarter and abnormal samples far from the center are removed, thereby improving the quality of training data.

[0107] (2) By performing SGMD decomposition on the power increment sequence, the trend disturbance in the original power sequence is weakened, and the model's ability to learn short-term fluctuations and local upslope changes is enhanced.

[0108] (3) By automatically optimizing the key hyperparameters of BiLSTM through WOA, the impact of manual parameter tuning on prediction performance is reduced.

[0109] (4) By using a conservative weight-bias adaptive correction layer, the magnitude weight and system bias of the main forecast output are corrected by quarter and time period to improve the model’s adaptability to seasonal and time period errors.

[0110] (5) By using error gating and conservative fusion mechanisms, excessive correction of the newly added correction layer is avoided, thereby improving the stability of complex model operation.

[0111] (6) The residual after weight-bias correction is learned by LSBoost to achieve secondary compensation for nonlinear residual.

[0112] (7) Improve the physical rationality and engineering usability of the prediction results by using peak-valley constraints, physical boundary constraints and final back-off fusion mechanism.

[0113] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0114] Figure 1 A general flowchart of a wind power prediction method according to an embodiment of the present invention is shown;

[0115] Figure 2 This diagram illustrates the scatter distribution of quarterly samples after fuzzy mapping and clustering, according to an embodiment of the present invention.

[0116] Figure 3 A heatmap of Pearson correlation analysis results according to an embodiment of the present invention is shown;

[0117] Figure 4 A heatmap of Spearman correlation analysis results according to an embodiment of the present invention is shown;

[0118] Figure 5 A comparison chart of the actual and predicted wind power values ​​before and after optimization is shown in an embodiment of the present invention.

[0119] Figure 6 A comparison chart of errors before and after optimization is shown in an embodiment of the present invention;

[0120] Figure 7 A comparison graph of training convergence curves before and after optimization is shown in an embodiment of the present invention. Detailed Implementation

[0121] To better understand the above-mentioned objects, features, and advantages of the present invention, such as Figures 1 to 7 As shown in the accompanying drawings and specific embodiments, the present invention will be further described in detail below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0122] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0123] Figure 1 Figure 1 shows the overall flowchart of a wind power prediction method according to an embodiment of the present invention. As shown in Figure 1, the method of the present invention includes the following steps in sequence according to the direction of the arrows: data acquisition and quarterly division, data preprocessing and feature construction, correlation analysis and single-center clustering screening, SGMD power increment decomposition, WOA-BiLSTM master prediction, conservative weight-bias correction, LSBoost residual compensation and constraint correction, and output of final prediction results and error evaluation. First, historical power data, meteorological data, and time information from wind farms were collected, and the samples were divided into different quarters based on the month. Then, missing value processing, outlier processing, and normalization were performed on the data, and features such as temperature, time, meteorological data, lagging power, and rolling statistics were constructed. Next, effective features were screened through correlation analysis, and outlier samples far from the center were removed using quarterly single-center clustering. Based on this, SGMD decomposition was performed on the power increment sequence to reduce the non-stationarity of the original sequence. Then, WOA was used to optimize the BiLSTM parameters, predict each modal component, and reconstruct the master prediction result. Next, a conservative weight-bias correction module was used to perform controlled correction of prediction biases for different quarters and time periods. Finally, LSBoost was used to compensate for the remaining residuals, and the final predicted value was output by combining physical boundaries, peak-valley constraints, and error gating mechanisms. As shown in Figure 1, the present invention forms a complete prediction process of "data purification - mode decomposition - optimized prediction - weight bias correction - residual compensation - constraint output". Among them, the conservative weight bias correction is located between the main prediction and residual compensation, and is a key link to improve the prediction accuracy and model stability in local time periods.

[0124] Figure 2 This diagram illustrates the scatter distribution of quarterly samples after fuzzy mapping and clustering, according to an embodiment of the present invention. Figure 2As shown in Figure 2, this figure illustrates the cluster scatter distribution of samples from the four quarters (Q1, Q2, Q3, and Q4) after fuzzy mapping, feature standardization, and PCA projection. Blue dots represent retained samples, red dots represent outliers, and yellow markers indicate the single-center cluster centers for each quarter. The scatter distribution shows that most samples are concentrated near the quarter center, indicating a relatively stable main distribution area for each quarter. However, some red outliers deviate significantly from the main sample group, indicating that these samples differ greatly from the normal operating state of that quarter. By setting a distance threshold from the quarter center, abnormal samples that deviate far from the center can be removed, reducing the interference of abnormal fluctuations, measurement errors, or extreme samples on model training. The conclusion drawn from Figure 2 is that quarterly single-center clustering screening can effectively identify and remove samples far from the quarterly main distribution, thereby improving the quality of training data and providing a more stable data foundation for subsequent SGMD decomposition and BiLSTM prediction.

[0125] Figure 3 A heatmap of Pearson correlation analysis results according to an embodiment of the present invention is shown. As shown in Figure 3, the... Figure 3 A heatmap showing the Pearson correlation analysis results between candidate input features and wind power is presented. The intensity of the colors in the heatmap reflects the strength of the linear correlation between different features; the closer the color is to the high correlation region, the stronger the linear correlation between the two variables; the closer the color is to the low correlation region, the weaker the linear correlation or the existence of a negative correlation. This graph allows observation of the linear relationship between meteorological features such as wind speed, maximum wind speed, air pressure, solar radiation intensity, temperature, rainfall, and wind direction and wind power, providing a basis for input feature selection and interpretation of feature importance. Figure 3 concludes that the degree of linear correlation between different meteorological variables and wind power varies. Features directly related to wind power changes generally have a high correlation, while some weakly correlated features contribute little to the model. Therefore, Pearson correlation analysis can be used to determine the linear effectiveness of candidate features and assist in constructing a more reasonable set of input features.

[0126] Figure 4 A heatmap of Spearman correlation analysis results according to an embodiment of the present invention is shown. As shown in Figure 4, the... Figure 4Figure 4 shows a heatmap of the Spearman correlation analysis results between each candidate input feature and wind power. Unlike Pearson correlation, which primarily reflects linear relationships, Spearman correlation measures monotonic relationships between variables, thus reflecting, to some extent, nonlinear but monotonic correlation characteristics. Different colors in the figure represent the strength of the monotonic correlation between features, which can be used to determine whether certain features, while not necessarily exhibiting a strictly linear relationship with wind power, may still show a stable monotonic trend. Figure 4 concludes that Spearman correlation analysis can supplement the shortcomings of Pearson correlation analysis, further evaluating the effectiveness of meteorological and power features from a monotonic correlation perspective. Combining Pearson and Spearman correlation analyses allows for a more comprehensive identification of input variables contributing to wind power prediction, improving the rationality of feature selection and model interpretability.

[0127] Figure 5 Figure 5 shows a comparison chart of the actual and predicted wind power values ​​before and after optimization according to an embodiment of the present invention. Figure 5 The figures show a comparison of the actual wind power output, single BiLSTM prediction, and optimized model prediction for representative days in Q1, Q2, Q3, and Q4. The actual values ​​in the figures reflect the actual changes in wind power output at different hourly points within a day. The single BiLSTM prediction is used as the baseline model, while the optimized model prediction is the final prediction result after SGMD decomposition, WOA parameter optimization, conservative weight-bias correction, and LSBoost residual compensation. The quarterly sub-figures show that wind power output exhibits significantly different fluctuation amplitudes and peak-valley characteristics across different quarters. The single BiLSTM model is prone to bias during some peak or trough periods, while the optimized model is closer to the actual value for most periods. Figure 5 concludes that compared to the single BiLSTM model, the optimized model can better track the trend of actual power output, especially showing better fitting ability during some peak, trough, and periods with significant fluctuations. This demonstrates that the proposed combined prediction and correction method can improve the accuracy of wind power output prediction.

[0128] Figure 6 A comparison chart of errors before and after optimization in an embodiment of the present invention is shown. As shown in Figure 6, the... Figure 6The absolute errors of the single BiLSTM model and the optimized model are compared for representative days across four quarters. A smaller error indicates a closer approximation between the predicted and actual values. Comparing the error bars at each hour reveals that the single BiLSTM model exhibits larger errors in certain time periods, particularly during power spikes, peak increases, or trough decreases. The optimized model, after SGMD decomposition, WOA parameter optimization, conservative weight-bias correction, and LSBoost residual compensation, shows a significant reduction in error for most time periods. Figure 6 concludes that the optimized model effectively reduces the prediction bias of the single BiLSTM model in certain time periods, demonstrating the positive effect of conservative weight-bias correction and residual compensation mechanisms in reducing localized high errors. Furthermore, this figure visually reflects the differences in error distribution across different quarters and time periods, providing a basis for further time-specific correction and model parameter optimization.

[0129] Figure 7 Figure 7 shows a comparison of training convergence curves before and after optimization in an embodiment of the present invention. Figure 7 The figure shows a comparison of the convergence curves of the single BiLSTM and the optimized model during the training process in four quarters: Q1, Q2, Q3, and Q4. The horizontal axis represents the number of iterations, and the vertical axis represents the training error. The curves for each quarter show that the training error of both the single BiLSTM and the optimized model decreases with the increase of the number of training iterations, indicating that the model can gradually learn the temporal patterns in the wind power sequence. Compared to the single BiLSTM, the error curve of the optimized model decreases faster overall or eventually stabilizes at a lower level, indicating that WOA parameter optimization, SGMD decomposition, and subsequent correction and compensation mechanisms help improve the model training effect. The conclusion drawn from Figure 7 is that the optimized model has better convergence and stability, and can more effectively reduce errors during training, demonstrating that the proposed method not only improves prediction accuracy but also enhances the reliability of the model training process.

[0130] The technical solution of the present invention will be illustrated below with two specific embodiments. Specific Implementation Example 1:

[0132] The wind power prediction method in this specific embodiment includes the following steps: data acquisition, correlation analysis, data preprocessing, quarterly single-center threshold screening, feature construction, SGMD decomposition, WOA-optimized BiLSTM prediction, master prediction fusion, conservative weight-bias adaptive correction, LSBoost residual compensation, and physical constraint and error gating output. Details are as follows:

[0133] (1) Data acquisition and correlation analysis:

[0134] Suppose the wind farm is at time... The actual power is The candidate meteorological feature vector is Before the features are included in the model, the Pearson correlation coefficient and Spearman correlation coefficient between the candidate features and wind power are calculated respectively.

[0135] ;

[0136] ;

[0137] In the formula, Describing covariance, Indicates standard deviation, This represents the rank transformation. Pearson correlation is used to reflect linear correlations, while Spearman correlation is used to reflect monotonic correlations. Correlation analysis is used in this invention as a means of feature validity analysis and model interpretation.

[0138] (2) Data preprocessing and quarterly division:

[0139] The original data underwent time format standardization, missing value handling, outlier / invalid value removal, and numerical transformation. When the original data did not provide quarterly labels, the samples were divided into Q1, Q2, Q3, and Q4 based on the month. Continuous features were normalized.

[0140] ;

[0141] In the formula, To prevent extremely small constants with a denominator of zero, quarterly divisions allow the model to learn the fluctuation patterns of wind power in different seasons.

[0142] (3) Quarterly single-center clustering threshold screening:

[0143] For each quarter Construct cluster screening feature matrix Preferably, This includes temperature fuzzy mapping features, power normalization features, temperature normalization features, power first-order difference normalization features, and moving standard deviation normalization features. For Standardization .

[0144] ;

[0145] Unlike traditional clustering screening methods with multiple cluster centers, this invention sets only one quarterly center within each quarter. , to represent the principal distribution center of the sample for that quarter:

[0146] ;

[0147] For quarter Inner Sample Calculate its value to the quarterly center Euclidean distance:

[0148] ;

[0149] Let the distance threshold be the distance set of this quarter. Quantiles, preferred :

[0150] ;

[0151] when When samples are retained, This method involves timely sample removal. It retains the main distribution samples within each quarter and removes outliers that are far from the quarter center, reducing the impact of outliers on subsequent model training.

[0152] (4) Feature construction and temperature fuzzy mapping:

[0153] This invention constructs multi-source enhancement features, including temperature fuzzy mapping features, time period features, hysteresis power features, differential features, sliding statistical features, rolling extremum features, and meteorological enhancement features. For temperature... A fuzzy mapping vector is constructed using three types of membership functions: low temperature, medium temperature, and high temperature.

[0154] ;

[0155] The time periodic feature is encoded using trigonometric functions:

[0156] ;

[0157] ;

[0158] in, For the hour, The week number is used. Lag power, sliding statistics, and rolling extrema are used to characterize the short-term inertia, fluctuation intensity, and local peak-valley state of power.

[0159] (5) SGMD power increment decomposition:

[0160] To reduce the trend and non-stationarity of the original power sequence, this invention does not directly address the issue. Instead of decomposing, we construct a sequence of power increments:

[0161] ;

[0162] Using SGMD Decomposed into One modal component:

[0163] ;

[0164] In the formula, For the first One modal component, To decompose the residual terms, the power increment sequence can be decomposed to highlight short-term ramps, local peaks and valleys, and sudden fluctuations, thus reducing the difficulty for BiLSTM to learn each modal component.

[0165] (6) WOA optimized BiLSTM master prediction model:

[0166] BiLSTM prediction models are established for each modal component obtained from SGMD decomposition. BiLSTM extracts bidirectional dependency features of the time series using two LSTM units, one forward and one backward.

[0167] ;

[0168] The hyperparameters of the BiLSTM were optimized using WOA, and the optimization variables were:

[0169] ;

[0170] In the formula, To hide the number of units, For learning rate, For L2 regularization parameters, For dropout parameters. The objective function of WOA is the weighted sum of the validation set RMSE and MAE:

[0171] ;

[0172] ;

[0173] ;

[0174] WOA searches for suitable parameters in the parameter space through a combination of encirclement, spiral update, and random search mechanisms. The minimum parameter combination is used to improve the prediction performance of BiLSTM.

[0175] (7) Master prediction fusion:

[0176] The prediction results for each modal component are reconstructed to obtain the prediction increment of SGMD-WOA-BiLSTM:

[0177] ;

[0178] And restore to the power prediction value:

[0179] ;

[0180] Simultaneously, the predicted values ​​of a single BiLSTM base model, hysteresis power, moving average, rolling maximum, and rolling minimum are all incorporated into a non-negative weighted fusion to obtain the main prediction result. :

[0181] ;

[0182] ;

[0183] in, Indicates forecasts or statistical characteristics from different sources. This indicates the corresponding fusion weight.

[0184] (8) Conservative weight-bias adaptive correction module:

[0185] One of the main improvements of this invention is the introduction of a conservative weight-bias adaptive correction layer after the master prediction result. This module does not directly replace BiLSTM, nor does it simply stack residuals. Instead, it learns the amplitude weight correction coefficients and bias correction terms of the master prediction output based on the historical errors of different quarters and different hour points in the training set.

[0186] set up Indicates quarter, For each hour, the principal prediction residual in the training set is:

[0187] ;

[0188] For each hour point The following linear correction relationship can be fitted using training samples:

[0189] [Quarterly-Period Linear Weighted Bias Fitting Formula]:

[0190] ;

[0191] in, This is the weighting correction coefficient, used to adjust the magnitude of the main prediction; This is a bias correction term used to correct systematic overestimation or underestimation. To avoid overfitting due to insufficient sample size, this invention employs a shrinkage estimation:

[0192] [Sample Size Adaptive Shrinkage Estimation Formula]:

[0193] ;

[0194] ;

[0195] ;

[0196] In the formula, This represents the number of training samples for that hour. For shrinkage strength parameters, To train the global median of the residuals. To further ensure stability, [the following is done]... and Set constraints:

[0197] ;

[0198] ;

[0199] in, and The 20th and 80th quantiles of the training residual distribution are preferably selected. Finally, a conservative fusion method is used to output the weight-bias correction result.

[0200] [Conservative Weighted Bias Fusion Correction Formula]:

[0201] ;

[0202] In the formula, A conservative fusion coefficient is preferred, ideally set at 0.30. This structure can subtly correct for amplitude deviations and center biases at different time periods without disrupting the overall trend of the main forecast.

[0203] (9) Error gating mechanism:

[0204] To avoid overcorrection caused by weight-bias correction, this invention incorporates a training error gating mechanism. Let the master prediction training error be... The training error after weight-bias correction is :

[0205] ;

[0206] ;

[0207] when When, retain the weight-bias correction result; when At that time, the main prediction result was regressed:

[0208]

Original Formula 9: Training Error Gating Formula

[0209] ;

[0210] This mechanism ensures that the newly added parameter correction module only takes effect when it can improve the training fit, thereby improving prediction stability.

[0211] (10) LSBoost residual compensation:

[0212] After weight-bias correction and error gating, an LSBoost residual compensation model is constructed. The residual learning objective is:

[0213] ;

[0214] The input features of LSBoost include Single BiLSTM predictions, lag power, moving mean, power difference, rolling maximum, rolling minimum, and meteorological features. LSBoost iteratively fits the residuals through multiple rounds of weak learner iterations.

[0215] ;

[0216] ;

[0217] In the formula, For residual compensation feature vectors, This is the quarterly residual scaling factor. To prevent excessive residual compensation, it can be... The training residual predictions are limited to the low and high quantile ranges.

[0218] (11) Fusion of peak-valley physical constraints and final backoff:

[0219] To ensure the physical feasibility of wind power prediction, this invention imposes upper and lower boundary constraints on the compensation results. Let the maximum allowable power be... ,but:

[0220] ;

[0221] Define the trough threshold, peak threshold, and super-peak threshold based on the training power distribution:

[0222] ;

[0223] To limit artificially high values ​​during off-peak periods and overshoot during peak and super-peak periods, the prediction results are made more consistent with the operating patterns of wind power. If the test error of the final optimized model is significantly greater than the error of a single BiLSTM benchmark, a backoff fusion is adopted.

[0224] [Final Error Backoff Fusion Formula]:

[0225] ;

[0226] in, The preferred value is 1.08. The preferred value is 0.82. This fallback mechanism is used to prevent the performance of complex models from degrading on individual test sets.

[0227] (12) Error evaluation index:

[0228] The present invention can use the following indicators to evaluate the prediction effect: root mean square error RMSE, mean absolute error MAE, mean deviation error MBE, maximum error MaxError, minimum error MinError, mean error MeanError, and standard error StdError.

[0229] ;

[0230] ;

[0231] ;

[0232] ;

[0233] MBE can be positive or negative. A positive value indicates that the overall valuation is overvalued, while a negative value indicates that the overall valuation is undervalued.

[0234] Table 1 Comparison of Error Evaluation Indicators Before and After Optimization

[0235] Q1 BILSTM 18.988 15.557 -6.5914 35.955 -80.592 -6.5914 18.981 Q1 SGMD-WOA-BILSTM 14.616 6.6299 1.1896 2.4698 -59.592 1.1896 13.066 Q2 BILSTM 7.8034 5.6937 -2.4401 16.381 -49.024 6.2269 -2.4401 Q2 SGMD-WOA-BILSTM 5.8968 2.6337 0.19502 11.402 -38.62 1.0261 0.195 Q3 BILSTM 10.4 7.4991 -2.2121 19.718 -40.713 -2.2121 10.178 Q3 SGMD-WOA-BILSTM 4.7809 2.0922 -1.5824 12.786 -40.29 -1.5824 4.5184 Q4 BILSTM 20.138 15.621 -5.3058 40.311 -77.093 -5.3058 20.104 Q4 SGMD-WOA-BILSTM 9.8905 6.0594 -1.6196 13.337 -64.431 -1.6196 8.3596

[0236] Table 1 shows a comparison of error evaluation indicators before and after optimization in this specific embodiment. As shown in Table 1, this table quantitatively compares the single BiLSTM and the optimized model using multiple indicators such as root mean square error (RMSE), mean absolute error (MAE), mean deviation error (MDE), maximum error, minimum error, mean error, and standard error. RMSE reflects the overall prediction error level, MAE measures the average deviation, MDE judges the overall overestimation or underestimation tendency of the model, maximum and minimum errors reflect extreme error situations, and standard error evaluates the degree of error fluctuation. The results for each quarter in the table show that the optimized model is generally lower than the single BiLSTM in key error indicators such as RMSE and MAE, indicating that the optimized model has higher prediction accuracy. Meanwhile, the improvement in MDE and standard error indicates that the optimized model can reduce systematic bias and improve the stability of error distribution. Figure 7 concludes that the proposed SGMD-WOA-BiLSTM-LSBoost method, combined with conservative weight-bias correction, exhibits better overall performance across multiple quarters and multiple error indicators, verifying the effectiveness of this method compared to the single BiLSTM.

[0237] In this specific embodiment, the present invention relates to the fields of new energy power generation prediction, short-term power prediction of wind farms, time series decomposition, deep learning parameter optimization, and ensemble learning residual correction technology, specifically to a wind power prediction method based on conservative weight-bias adaptive correction (SGMD-WOA-BiLSTM-LSBoost). This method acquires historical power data, meteorological data, and time information from wind farms, and sequentially completes quarterly division, correlation analysis, data preprocessing, feature construction, and single-center clustering threshold screening. SGMD decomposition is performed on the power increment sequence, and WOA is used to optimize the number of hidden units, learning rate, regularization parameter, and dropout parameter of the BiLSTM network. The optimized BiLSTM is then used to predict and reconstruct the main prediction result for each modal component. To improve the adaptability of the master forecasting model to seasonal and temporal biases, this invention introduces a conservative weight-bias adaptive correction layer after the master forecast output. Based on the distribution of the training residuals across quarterly and hourly dimensions, it learns the weight correction coefficients and bias correction terms, and limits over-correction through shrinkage estimation, parameter limiting, conservative fusion, and error gating mechanisms. Finally, LSBoost is used to perform secondary compensation on the corrected residuals, and the final wind power forecast value is output by combining peak-valley constraints and physical boundary constraints. This specific embodiment can improve the accuracy, stability, and engineering usability of wind power forecasting, and is particularly suitable for wind power forecasting scenarios with seasonality, short-term fluctuations, and local peak-valley abrupt changes. Specific Implementation Example 2:

[0239] The wind power prediction method in this specific embodiment two includes:

[0240] (1) Data preparation:

[0241] The system reads time-series wind power data and corresponding meteorological data, automatically identifying the time, temperature, and power columns. It also identifies meteorological variables such as wind speed, wind direction, rainfall, maximum wind speed, air pressure, and solar radiation intensity based on available fields. If the data lacks a quarter field, it automatically generates quarter labels based on the month. Invalid time, power, and temperature samples are deleted, and available meteorological features are numerically converted.

[0242] (2) Quarterly single-center cluster screening:

[0243] A clustering feature matrix is ​​constructed for each quarter, with columns including temperature fuzzy mapping, power normalization, temperature normalization, power difference normalization, and power moving standard deviation normalization. After standardizing this matrix, the mean of all samples within the quarter is used as a single center. The Euclidean distance from each sample to this center is calculated, and the farthest sample is removed using the 99th percentile of the distance as a threshold. This step does not aim to divide the quarterly samples into multiple categories, but rather uses the quarterly principal distribution center as a benchmark for filtering outliers.

[0244] (3) Master model prediction:

[0245] With power increment sequence As the object of SGMD decomposition, multiple modal components are obtained. WOA is used to search for the number of hidden units, learning rate, regularization parameters, and dropout parameters of BiLSTM, and a validation set is used. As the objective function, a BiLSTM prediction model is trained for each modal component. The modal prediction results are summed and reconstructed into a power increment prediction, which is then added to the power at the previous time step to obtain the SGMD-WOA-BiLSTM prediction value.

[0246] (4) Conservative weighting-bias correction:

[0247] After master prediction fusion, fitting is performed for each quarter and each hour based on the training samples. If the sample size is small in a particular hour, the global bias median is used, and then... Shrink the parameters. Ultimately, Limiting it to 0.95 to 1.05 will Limited to the 20% to 80% quantile range of the training residual distribution, and with Perform conservative fusion. If the training error does not decrease after correction, revert to the original master prediction result.

[0248] (5) LSBoost residual compensation and output:

[0249] After weight-bias correction, calculate the residuals. ,by The base model's predicted values, lag power, moving mean, difference, rolling extreme values, and meteorological features are used as inputs to train the LSBoost residual model. During the testing phase, the predicted residuals are used. and superimpose it onto Then, nonnegative constraints, maximum power constraints, peak-valley constraints, and error backoff are applied and fused to obtain the final wind power prediction value.

[0250] Based on specific embodiments one and two, the innovative points of this invention can be summarized as follows:

[0251] (1) A conservative weight-bias adaptive correction method is proposed:

[0252] Existing BiLSTM wind power prediction methods typically rely on backpropagation to update the weights within the network, or on intelligent optimization algorithms to adjust global hyperparameters such as the number of hidden layer units, learning rate, and regularization coefficient. These methods primarily focus on optimizing the overall model parameters and struggle to provide fine-grained corrections for local overestimations, underestimations, and peak-valley biases that may occur in different quarters or time periods.

[0253] Therefore, this invention constructs quarterly and time-period-specific weighted correction coefficients and bias correction terms after the SGMD-WOA-BiLSTM master forecast result, performing secondary parameterization correction on the master forecast result. Its basic correction form is as follows:

[0254]

Original Formula 1

[0255] ;

[0256] in, Indicates the first The quarter, the The time period, the first The principal prediction power for each sample; This represents the predicted power after weight-bias correction; Indicates the first The quarter, the The weighting correction coefficients for each time period are used to adjust the prediction amplitude. Indicates the first The quarter, the The bias correction term for each time period is used to correct for systematic over- or under-biasing.

[0257] This method can modify the originally uniform global forecast output into a local parameterized output that is seasonal and time-specific, thereby improving the model's ability to adapt to seasonal and time-specific errors.

[0258] (2) A controlled correction mechanism with conservative fusion coefficients is proposed:

[0259] To avoid excessive weight and bias correction causing the prediction results to deviate from the original BiLSTM master prediction trend, this invention further introduces a conservative fusion coefficient to fuse the weight-bias correction results with the original master prediction results, instead of directly replacing the original prediction values ​​with the correction results.

[0260] Its fusion form is as follows:

[0261]

Original Formula 3

[0262] ;

[0263] Substituting the basic correction formula, we get:

[0264] ;

[0265] in, This represents the predicted power after conservative weighting and bias correction. Denotes the conservative fusion coefficient, and satisfies Preferred to satisfy .

[0266] This design allows the correction module to make only small, controlled corrections to the main prediction results. It can adjust local biases using historical residuals while preserving the overall trend learned by the BiLSTM main model, thus avoiding overly aggressive corrections that could amplify errors.

[0267] (3) A weight and bias learning mechanism under amplitude constraint is proposed:

[0268] This invention does not directly use training residuals to forcibly correct prediction results. Instead, it learns the original weight coefficients and bias terms based on the error distribution of different quarters and time periods in the training samples, and constrains them through a limiting function, thereby improving the stability of the correction process. The training residuals are defined as:

[0269] ;

[0270] The weights and biases can be obtained by fitting the training samples from that quarter and that period:

[0271] ;

[0272] The final correction parameters are then obtained through the limiting function:

[0273]

Original Formula 2

[0274] ;

[0275] ;

[0276] in, This represents the original weight correction coefficients obtained from the fitting process; This represents the original bias correction term obtained from the fitting; Indicates will Limited to arrive Within the range. Preferably, the weight correction coefficient satisfies .

[0277] This limiting mechanism can prevent a small number of abnormal samples from causing the predicted value to be over-amplified or under-amplified for a certain period of time, making the weight-bias correction more robust.

[0278] (4) A hierarchical error correction structure of "weight-bias correction + LSBoost residual compensation" is proposed:

[0279] Existing residual compensation methods typically learn directly from the master prediction residuals, causing the residual model to simultaneously undertake the tasks of systematic bias correction and nonlinear error fitting, which can easily lead to overly strong residual compensation or local overfitting.

[0280] This invention divides the error correction process into two layers: the first layer corrects seasonal and periodic regular deviations through conservative weight-bias correction; the second layer then uses LSBoost to learn and compensate for the remaining nonlinear residuals. The residual after conservative weight-bias correction is defined as:

[0281] ;

[0282] Learning residual mappings using LSBoost:

[0283] ;

[0284] The final prediction result is expressed as follows:

[0285]

Original Formula 4

[0286] ;

[0287] in, This represents the corrected master prediction residual; This represents the input features for residual compensation; This represents the LSBoost residual mapping model; This represents the residual compensation value obtained from LSBoost prediction; Indicates the first The residual scaling factor corresponding to each quarter; This represents the final predicted wind power output.

[0288] Through this hierarchical structure, conservative weight-bias correction is responsible for correcting regularity biases, and LSBoost is responsible for compensating for the remaining complex residuals, thereby reducing the learning pressure on the residual model and improving the stability and robustness of the prediction results.

[0289] (5) A local parameterization correction method for seasonal and temporal errors in wind power is proposed:

[0290] Wind power output is significantly affected by seasons, weather, and time of day, with inconsistent error distributions across different quarters and hours. Traditional global parameter tuning methods struggle to address these varying errors effectively.

[0291] The present invention sets the correction parameters to and This ensures that each quarter and each time period has independent amplitude correction parameters and bias correction parameters. That is, different correction relationships are used for different quarters and different hourly points.

[0292] ;

[0293] For example, if the midday peak forecast is too low, it can be addressed by appropriately increasing... or The forecast value was revised upwards; for cases where the forecast was too high during the nighttime off-peak period, it could be adjusted by reducing the forecast value. Or set a negative The forecast value has been lowered.

[0294] This method is more refined than single global parameter correction, and more adaptive and interpretable than manual empirical point-in-time correction. It can better adapt to the seasonal, temporal, and local peak-valley variation characteristics of wind power sequences.

[0295] In summary, the purpose of this invention is to provide a wind power prediction method based on conservative weight-bias adaptive correction (SGMD-WOA-BiLSTM-LSBoost). This method improves sample quality through quarterly single-center threshold screening, reduces the non-stationarity of power increment sequences through SGMD, optimizes BiLSTM network parameters through WOA, corrects quarterly and periodic output biases through a conservative weight-bias correction layer, and performs secondary compensation for the remaining residuals through LSBoost. This improves the accuracy, robustness, and engineering usability of wind power prediction.

[0296] Existing BiLSTM wind power prediction methods typically update the network's internal weights through backpropagation or adjust global hyperparameters such as the number of hidden layer units, learning rate, and regularization coefficients through intelligent optimization algorithms. However, these methods struggle to finely correct for local overestimations, underestimations, and peak-valley deviations that exist in different quarters and time periods. To address this, this invention proposes a quarterly and time-period-specific adaptive weight-bias correction method. After the SGMD-WOA-BiLSTM master prediction result is output, a weight correction coefficient α(q,h) and a bias correction term β(q,h) are constructed based on the historical prediction residuals from different quarters and hourly points in the training set.

[0297] pass The parameter λ = 0.30 is used to correct the main prediction result. This method can simultaneously adjust the prediction amplitude and system offset, enabling the model to better adapt to the seasonal differences and time-specific error distributions of wind power. Subsequently, the corrected prediction result is input into the LSBoost residual compensation model to perform secondary learning on the remaining nonlinear residuals, thereby improving prediction accuracy and robustness.

[0298] This invention introduces a conservative weight-bias adaptive correction layer between the SGMD-WOA-BiLSTM master forecast and LSBoost residual compensation, enabling explicit correction of the BiLSTM output amplitude and system bias. Error gating and backoff fusion mechanisms ensure the stability and reliability of the correction process. This method is suitable for short-term or ultra-short-term wind power forecasting, especially for wind farm data exhibiting significant seasonality, local peak-valley characteristics, and short-term random fluctuations.

[0299] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting wind power output, characterized in that, include: Step S1: Obtain samples and divide the samples into Q1, Q2, Q3 and Q4 according to quarters; wherein, the samples include: historical wind power data of wind farms, meteorological data and time information; Step S2: Perform time standardization, missing value handling, outlier handling, normalization, and effective sample screening on the samples; Step S3: Construct an input feature set; wherein the input feature set includes: temperature fuzzy mapping features, time period features, hysteresis power features, difference features, moving mean features, moving standard deviation features, rolling extreme value features, and meteorological enhancement features; Step S4: Construct a clustering feature matrix for the samples of each quarter, and use a single-center clustering screening method to determine the quarter center. Then, based on the distance threshold from the samples of each quarter to the quarter center, remove abnormal samples that are far from the center. Step S5: Perform SGMD decomposition on the power increment sequence to obtain multiple modal components; Step S6: Optimize the BiLSTM network using WOA, and use the optimized BiLSTM to predict each modal component separately to obtain the prediction results for each modal component; Step S7: Reconstruct the prediction results of each modal component to obtain the SGMD-WOA-BiLSTM master prediction result; Step S8: Based on the master prediction residuals of different quarters and different time periods in the training set, learn the corresponding weight correction coefficients and bias correction terms, and construct a conservative weight-bias correction model to achieve parameterized correction of the master prediction results. Step S9: Using the residual between the parameterized corrected prediction result and the actual power as the learning object, construct the LSBoost residual compensation model to achieve secondary compensation of the prediction result after conservative weight-bias correction. Step S10: Perform physical boundary constraints, peak-valley constraints and error gating on the compensated prediction results, and output the final wind power prediction value.

2. The wind power prediction method according to claim 1, characterized in that, In step S3, the temperature fuzzy mapping feature is represented by three membership functions: low temperature, medium temperature, and high temperature. The calculation method is as follows: ； In the formula, express Temperature value at time, Indicates the membership degree at low temperature. Indicates the degree of membership at medium temperature. Indicates the membership degree at high temperature. express Temperature fuzzy mapping characteristics at any given time; The time period characteristics include: hourly period characteristics and weekly period characteristics, calculated as follows: ； ； ； ； In the formula, express The hour corresponding to the time. express The week number corresponding to the time. Represents hourly sine code. Represents hourly cosine encoding. Represents the sine code for the day of the week. Represents the cosine encoding of the weekday; The hysteresis power characteristics include , , , and At least one of the following; the difference feature includes first-order difference. and second-order difference At least one of the following: the moving mean feature, the moving standard deviation feature, and the rolling extreme value feature are used to characterize the average level, fluctuation intensity, and local peak and valley state of the power sequence within a preset time window, respectively; the meteorological enhancement feature includes at least one of wind speed, wind direction, rainfall, maximum wind speed, air pressure, and solar radiation intensity.

3. The wind power prediction method according to claim 1, characterized in that, In step S4, the single-center clustering screening method is performed quarterly, for the first quarter... A cluster feature matrix is ​​constructed from quarterly samples. and calculate the quarterly center ; The quarterly center The calculation method is as follows: ； For the first For each sample, calculate its Euclidean distance to the quarter center: ； Based on the distance distribution, the quarterly distance threshold is determined as follows: ； In the formula, Indicates the first Clustering feature matrix for each quarter, Indicates the first Single cluster center for each quarter Indicates the first The number of samples participating in clustering within a quarter. Indicates the first Euclidean distance from each sample to the quarter center Indicates the first Distance threshold for each quarter, Indicates the quantile threshold; when When, retain the sample; when If the sample is found to be an anomaly far from the quarterly center, it will be removed.

4. The wind power prediction method according to claim 1, characterized in that, In step S5, the decomposition object of SGMD is the power increment sequence, which is represented as: ； In the formula, express Real-time wind power output express Real-time wind power output express Power increment at any moment; The power increment sequence was decomposed using SGMD to obtain multiple modal components: ； In the formula, Indicates the first One modal component, Indicates the number of modal components. This indicates the decomposition of residual terms.

5. The wind power prediction method according to claim 1, characterized in that, In step S6, WOA is used to optimize the parameters of the BiLSTM network, and the parameter vector to be optimized is represented as follows: ； In the formula, Indicates the number of hidden layer units in the BiLSTM. This represents the initial learning rate. Represents the L2 regularization coefficient. Indicates the dropout ratio; The objective function of WOA uses a weighted combination of the root mean square error and the mean absolute error of the validation set, expressed as: ； In the formula: ； ； In the formula, ; WOA minimizes To obtain the optimal parameters: ； In the formula, This represents the optimal parameter combination for BiLSTM obtained through WOA optimization; For each SGMD modal component, the optimized BiLSTM is used for prediction, resulting in: ； In the formula, Indicates the first Each modal component in Predicted value at time, express Input feature sequence at any time, This represents the BiLSTM parameters optimized by WOA.

6. The wind power prediction method according to claim 1, characterized in that, In step S7, the prediction results of each modal component are reconstructed into predicted power increment values: ； Further, the master predicted power is obtained: ； Alternatively, the main prediction power can be obtained through non-negative weighted fusion: ； In the formula, This represents the BiLSTM modal reconstruction prediction value. Indicates the hysteresis power characteristic. Indicates the characteristics of the moving average. Indicates the feature of maximum scrolling value. Indicates the rolling minimum value characteristic; , , , and As weight; Weight , , , and satisfy: ； 。 7. The wind power prediction method according to claim 1, characterized in that, In step S8, the conservative weight-bias correction model is constructed separately for each quarter and time period, for the first... The quarter, the The master forecast results for each time period are corrected and expressed as follows: ； In the formula, Indicates the first The quarter, the The time period, the first The principal prediction power for each sample This represents the predicted power after weight-bias correction. Indicates the first The quarter, the Weighting correction coefficients for each time period Indicates the first The quarter, the Bias correction term for each time period; The training residual is represented as: ； In the formula, Indicates the first The quarter, the The time period, the first The actual wind power output of a sample This represents the master prediction residual for the corresponding sample; The weight correction coefficients and bias correction terms are obtained by fitting the training samples for that quarter and that period: ； The final weight correction coefficients and bias correction terms are obtained through the amplitude limiting function: ； ； In the formula, Indicates will Limited to Within the range, and These represent the lower and upper limits of the weight correction coefficient, respectively. and These represent the lower and upper limits of the bias correction term, respectively.

8. The wind power prediction method according to claim 7, characterized in that, The conservative weight-bias correction model further introduces a conservative fusion coefficient. This merges the correction result with the original master prediction result, and is expressed as: ； Right now: ； In the formula, This represents the predicted power after conservative weighting and bias correction. Denotes the conservative fusion coefficient, and .

9. The wind power prediction method according to claim 1, characterized in that, In step S9, the LSBoost residual compensation model constructs residuals based on the prediction results after conservative weight-bias correction, and these residuals are expressed as follows: ； Constructing residual compensation input features: ； In the formula, This represents the prediction value of a single BiLSTM base model. Indicates the hysteresis power characteristic, Indicates the characteristics of the moving average. Indicates differential power characteristics, Indicates the maximum scrolling value characteristic. Indicates the rolling minimum value characteristic. Represents a set of meteorological enhancement features; Learning residual mappings using LSBoost: ； The final compensated prediction value is: ； In the formula, Indicates the first The residual scaling factor for each quarter This represents the residual compensation value obtained from LSBoost prediction.

10. The wind power prediction method according to claim 1, characterized in that, In step S10, the peak-valley constraint includes: valley constraint, peak constraint, and super-peak constraint; The physical boundary constraints are expressed as follows: ； In the formula, This indicates the reasonable upper limit of wind power output; The trough constraint, peak constraint, and superpeak constraint are respectively represented as follows: ； ； ； In the formula, Indicates the trough threshold. Indicates the peak threshold. Indicates the peak value threshold. Indicates the upper limit of the trough. Indicates the upper bound of the peak value. Indicates the upper bound of the superpeak value; The error gating process is expressed as follows: ； In the formula, This represents the root mean square error of the optimization model. This represents the root mean square error of a single BiLSTM basis model. Indicates the error gating threshold. This represents the fallback fusion coefficient.

Images