Wind farm day-ahead wind power prediction method and system
By constructing a robust auxiliary classification generative adversarial network (RAC-GAN) and a clustering algorithm, the problems of model complexity and inaccurate prediction in wind power forecasting are solved, and more accurate wind power forecasting is achieved.
Patent Information
- Application Number
- CN202111270931.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-29
- Publication Date
- 2026-02-24
- Estimated Expiration
- 2041-10-29
AI Technical Summary
Existing physics-based wind power prediction methods cannot extract inherent patterns from multi-source, multi-dimensional, and multi-modal data, resulting in inaccurate prediction results and complex model structures and cumbersome calculation processes.
A method for predicting day-ahead wind power in wind farms based on robust auxiliary classification generative adversarial network (RAC-GAN) is constructed. The original dataset is constructed using Pearson correlation coefficient, and k-means clustering and scene generation are performed. Scenes with high similarity to the day-ahead wind power are selected to generate point prediction and interval prediction results.
It improves the accuracy of point and interval prediction of wind power, and is more accurate and effective than traditional methods, maintaining good prediction results even in the presence of numerical weather forecast errors.
Smart Images
Figure CN114021799B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power generation technology, and in particular to a method and system for predicting day-ahead wind power in wind farms. Background Technology
[0002] In recent years, renewable energy sources, such as wind power, have received increasing attention. Unlike traditional power generation, wind power is characterized by complexity and instability. As the installed capacity of wind turbines continues to increase, the strong randomness and volatility of its output have posed challenges to the safe and stable operation of the power system. Therefore, accurate prediction of wind power output is extremely important for optimizing grid dispatch and improving wind farm efficiency. The current technical solution for wind power prediction based on physical methods is as follows: A combined weighted wind power prediction model is established, taking into account meteorological factors and the spatial correlation of wind speed in numerical weather prediction. Considering the low accuracy of wind speed in numerical weather prediction data, a numerical weather prediction wind speed correction model based on Gaussian processes is established, incorporating other meteorological factors such as wind direction, temperature, humidity, and air pressure for wind power prediction. Simultaneously, based on the spatial correlation analysis of wind speed between the target wind farm and adjacent wind farm areas, the delay time of the point with the maximum correlation coefficient is calculated, and a wind speed spatial correlation prediction model is established. Based on the wind power prediction model with numerical weather prediction bias correction and the spatial correlation prediction model, a combined weighted prediction model is established, and the weighting value of each individual model in the combined model is obtained using the Lagrange multiplier method to obtain the predicted wind power.
[0003] However, physics-based wind power prediction cannot find the inherent patterns in multi-source, multi-dimensional, and multi-modal data, nor can it uncover the relationships between historical data, resulting in inaccurate prediction results. Furthermore, physics-based wind power prediction models are complex in structure and involve cumbersome calculations. Summary of the Invention
[0004] To address the issues of insufficient accuracy, complex prediction model structure, and cumbersome calculation process in existing physics-based wind power prediction methods, this invention provides a day-ahead wind power prediction method for wind farms, comprising:
[0005] Based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power, an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted is constructed.
[0006] The predicted daily wind power is removed from the original dataset, resulting in a clustered dataset. K-means clustering is then performed to obtain a labeled original dataset. A robust auxiliary classification-based adversarial network is then used to generate a massive number of labeled scenes. The specific steps for generating a massive number of labeled scenes using the robust auxiliary classification-based adversarial network include:
[0007] A data encoder is introduced at the generator input of the auxiliary classification generative adversarial network, and shallow raw data features are pre-learned based on real wind power and its impact characteristics to obtain random noise input.
[0008] The generator generates a massive amount of sample data that satisfies the probability distribution characteristics of real samples based on the random noise.
[0009] The discriminator judges the quality of the generated samples based on the generated sample data and the original sample data, thereby expanding the training sample data;
[0010] The robust auxiliary classification generative adversarial network iterates through a game-theoretic optimization process to reduce the impact of noise, and finally generates a massive number of labeled wind power and its impact characteristics under noise interference.
[0011] Based on known historical daily wind power and numerical weather forecast meteorological characteristics, the cluster label category of the day to be predicted is determined. Then, based on the cluster label category, multiple scenarios with high similarity to the wind power of the day to be predicted are selected from the massive number of labeled scenarios to form a similar scenario set. The specific steps of selecting multiple scenarios with high similarity to the wind power of the day to be predicted based on the cluster label category to form a similar scenario set include:
[0012] Based on the cluster label of the predicted date, the corresponding generated sample is found in the massive labeled scenario;
[0013] Using the Pearson correlation coefficient between the predicted daily wind power and the influencing features as the weighting coefficient, the weighted PCC of all influencing features of the predicted daily wind power and the corresponding features of the j-th generated sample is set as PCC. j The expression is:
[0014]
[0015] Where PCC(x,y) i ) represents the PCC between the predicted daily wind power and the influencing characteristic i; PCC(y) i ,y ji () represents the PCC between known influence feature i and influence feature i of the j-th generated sample; PCC j The weighted PCC between the j-th generated sample and the features of the day to be predicted;
[0016] According to the PCC j From largest to smallest, select multiple samples with high historical daytime similarity that are highly correlated with the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set;
[0017] Based on the average, upper limit, and lower limit of the wind power on the day to be predicted in the similar scenario set, point prediction and interval prediction results of wind power at multiple set times on the day to be predicted are obtained respectively.
[0018] The present invention also provides a day-ahead wind power prediction system for wind farms, comprising:
[0019] The construction module is used to construct an original dataset containing the numerical weather forecast meteorological characteristics and the wind power of the day to be predicted, based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power.
[0020] The clustering generation module is used to remove the predicted daily wind power from the original dataset, obtain a clustered dataset, perform k-means clustering, obtain an original dataset with cluster labels, and generate an adversarial network based on robust auxiliary classification to generate a large number of labeled scenes.
[0021] The filtering module is used to determine the cluster label category of the day to be predicted based on the known historical daily wind power and the meteorological characteristics of numerical weather forecasts, and to filter out multiple scenarios with high similarity to the wind power of the day to be predicted from the massive number of labeled scenarios based on the cluster label category, forming a similar scenario set;
[0022] The prediction module is used to obtain point prediction and interval prediction results of wind power at multiple set times on the day to be predicted based on the average value, upper limit value and lower limit value of wind power on the day to be predicted in the similar scenario set.
[0023] The clustering generation module includes:
[0024] Clustering dataset unit is used to remove the predicted daily wind power from the original dataset to obtain the normalized features corresponding to the influencing features, thus forming a clustering dataset;
[0025] Clustering unit, used to perform k-means clustering on the data in the clustered dataset to obtain multi-cluster labels;
[0026] A labeling unit is used to assign the label to each scene in the original dataset.
[0027] The learning unit is used to introduce a data encoder at the generator input of the auxiliary classification generative adversarial network, and to pre-learn shallow raw data features based on real wind power and its impact characteristics to obtain random noise input.
[0028] The generator unit is used to generate massive amounts of sample data that satisfy the probability distribution characteristics of real samples based on the random noise.
[0029] The discriminator unit is used to determine the quality of the generated samples based on the sample data generated by the generator unit and the original sample data, thereby expanding the training sample data.
[0030] The optimized generation unit is used to assist the classification generative adversarial network in the iterative process, and to carry out game optimization in the direction of reducing the impact of noise, and finally generate a large number of labeled wind power and its impact characteristics under noise interference.
[0031] The filtering module includes:
[0032] The search unit is used to find the corresponding generated sample in the massive labeled scenario based on the cluster label of the predicted day.
[0033] The setting unit is used to set the weighted PCC of all influence features of the wind power to be predicted on the day of wind power and the influence features as the weighting coefficients, and the corresponding features between the j-th generated sample and the Pearson correlation coefficient between the wind power to be predicted on the day of wind power. j ;
[0034] Similar scene set unit, used according to the PCC j From largest to smallest, select multiple samples with high inter-day similarity to the historical data of the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set.
[0035] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, performs the following steps:
[0036] Based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power, an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted is constructed.
[0037] The predicted daily wind power is removed from the original dataset, resulting in a clustered dataset. K-means clustering is then performed to obtain a labeled original dataset. A robust auxiliary classification-based adversarial network is then used to generate a massive number of labeled scenes. The specific steps for generating a massive number of labeled scenes using the robust auxiliary classification-based adversarial network include:
[0038] A data encoder is introduced at the generator input of the auxiliary classification generative adversarial network, and shallow raw data features are pre-learned based on real wind power and its impact characteristics to obtain random noise input.
[0039] The generator generates a massive amount of sample data that satisfies the probability distribution characteristics of real samples based on the random noise.
[0040] The discriminator judges the quality of the generated samples based on the generated sample data and the original sample data, thereby expanding the training sample data;
[0041] The robust auxiliary classification generative adversarial network iterates through a game-theoretic optimization process to reduce the impact of noise, and finally generates a massive number of labeled wind power and its impact characteristics under noise interference.
[0042] Based on known historical daily wind power and numerical weather forecast meteorological characteristics, the cluster label category of the day to be predicted is determined. Then, based on the cluster label category, multiple scenarios with high similarity to the wind power of the day to be predicted are selected from the massive number of labeled scenarios to form a similar scenario set. The specific steps of selecting multiple scenarios with high similarity to the wind power of the day to be predicted based on the cluster label category to form a similar scenario set include:
[0043] Based on the cluster label of the predicted date, the corresponding generated sample is found in the massive labeled scenario;
[0044] Using the Pearson correlation coefficient between the predicted daily wind power and the influencing features as the weighting coefficient, the weighted PCC of all influencing features of the predicted daily wind power and the corresponding features of the j-th generated sample is set as PCC. j The expression is:
[0045]
[0046] Where PCC(x,y) i ) represents the PCC between the predicted daily wind power and the influencing characteristic i; PCC(y) i ,y ji () represents the PCC between known influence feature i and influence feature i of the j-th generated sample; PCC j The weighted PCC between the j-th generated sample and the features of the day to be predicted;
[0047] According to the PCC j From largest to smallest, select multiple samples with high historical daytime similarity that are highly correlated with the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set;
[0048] Based on the average, upper limit, and lower limit of the wind power on the day to be predicted in the similar scenario set, point prediction and interval prediction results of wind power at multiple set times on the day to be predicted are obtained respectively.
[0049] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the following steps:
[0050] Based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power, an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted is constructed.
[0051] The predicted daily wind power is removed from the original dataset, resulting in a clustered dataset. K-means clustering is then performed to obtain a labeled original dataset. A robust auxiliary classification-based adversarial network is then used to generate a massive number of labeled scenes. The specific steps for generating a massive number of labeled scenes using the robust auxiliary classification-based adversarial network include:
[0052] A data encoder is introduced at the generator input of the auxiliary classification generative adversarial network, and shallow raw data features are pre-learned based on real wind power and its impact characteristics to obtain random noise input.
[0053] The generator generates a massive amount of sample data that satisfies the probability distribution characteristics of real samples based on the random noise.
[0054] The discriminator judges the quality of the generated samples based on the generated sample data and the original sample data, thereby expanding the training sample data;
[0055] The robust auxiliary classification generative adversarial network iterates through a game-theoretic optimization process to reduce the impact of noise, and finally generates a massive number of labeled wind power and its impact characteristics under noise interference.
[0056] Based on known historical daily wind power and numerical weather forecast meteorological characteristics, the cluster label category of the day to be predicted is determined. Then, based on the cluster label category, multiple scenarios with high similarity to the wind power of the day to be predicted are selected from the massive number of labeled scenarios to form a similar scenario set. The specific steps of selecting multiple scenarios with high similarity to the wind power of the day to be predicted based on the cluster label category to form a similar scenario set include:
[0057] Based on the cluster label of the predicted date, the corresponding generated sample is found in the massive labeled scenario;
[0058] Using the Pearson correlation coefficient between the predicted daily wind power and the influencing features as the weighting coefficient, the weighted PCC of all influencing features of the predicted daily wind power and the corresponding features of the j-th generated sample is set as PCC. j The expression is:
[0059]
[0060] Where PCC(x,y) i) represents the PCC between the predicted daily wind power and the influencing characteristic i; PCC(y) i ,y ji () represents the PCC between known influence feature i and influence feature i of the j-th generated sample; PCC j The weighted PCC between the j-th generated sample and the features of the day to be predicted;
[0061] According to the PCC j From largest to smallest, select multiple samples with high historical daytime similarity that are highly correlated with the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set;
[0062] Based on the average, upper limit, and lower limit of the wind power on the day to be predicted in the similar scenario set, point prediction and interval prediction results of wind power at multiple set times on the day to be predicted are obtained respectively.
[0063] The wind farm day-ahead wind power prediction method and system provided by this invention constructs an original dataset containing the wind power of the day to be predicted, meteorological characteristics of numerical weather forecasts, and historical wind power through Pearson correlation coefficient. Based on clustering algorithm, a labeled original dataset is constructed. A scene generation model based on an improved robust auxiliary classification generative adversarial network is designed to generate massive labeled samples. Based on the known historical wind power and meteorological characteristics of numerical weather forecasts, the cluster label category of the day to be predicted can be determined. Scenes with high similarity to the wind power of the day to be predicted are selected under this cluster label, thereby obtaining the point prediction and interval prediction results of the wind power of the day to be predicted. This improves the accuracy of point prediction and interval prediction of wind power. Compared with the traditional WGP interval prediction method, the prediction results of this invention are more accurate and effective. Attached Figure Description
[0064] Figure 1 This is a schematic diagram illustrating the correlation between wind power at time t on the day to be predicted and various influencing characteristics in an embodiment of the present invention.
[0065] Figure 2 This is a sample format of the original dataset containing the daily wind power to be predicted and its impact characteristics in the embodiments of the present invention;
[0066] Figure 3 This is a schematic diagram of the RAC-GAN model principle provided in the embodiments of the present invention;
[0067] Figure 4 This is a schematic diagram of the wind farm day-ahead wind power prediction process provided in an embodiment of the present invention;
[0068] Figure 5 This is a flowchart of the wind farm day-ahead wind power prediction method provided in an embodiment of the present invention;
[0069] Figure 6 This is a schematic diagram comparing the wind power prediction results and prediction indicators of the method of the present invention and the WGP method under different confidence levels in the four seasons.
[0070] Figure 7 This is a schematic diagram of the wind farm day-ahead wind power prediction system provided in an embodiment of the present invention;
[0071] Figure 8 This is a schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention. Detailed Implementation
[0072] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0073] See Figure 4 and Figure 5 The wind farm day-ahead wind power prediction method provided in this embodiment of the invention specifically includes the following steps:
[0074] Step S101: Based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power, construct an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted.
[0075] In practical applications, wind power is affected by wind speed, wind direction, temperature, humidity, pressure, and historical daily wind power. The Pearson correlation coefficient (PCC) can be used to analyze minute local morphological differences without the need to normalize the wind power data, and can effectively analyze the correlation between wind power and various influencing characteristics. The PCC calculation formula is shown in the following formula (1):
[0076]
[0077] In equation (1), x represents the daily wind power to be predicted, and y represents the wind power to be predicted. i The influence characteristics of the predicted daily wind power output; x k y ik x and y are respectively i The k-th data point; x and y are respectively i The average value of the data; k and n are positive integers.
[0078] Wind speed, wind direction, temperature, humidity, and pressure are selected as the first preliminary features for predicting the wind power of the day to be predicted. Simultaneously, considering the influence of historical wind power on the predicted day's wind power, the correlation between the wind power of each historical day and the predicted day is analyzed based on the Pearson correlation coefficient. Historical days with a Pearson correlation coefficient absolute value greater than a preset threshold are selected as the second preliminary features. For example, in this embodiment, wind power within 5 historical days is selected as the second preliminary feature. These 10 selected preliminary features are set as the influencing features of the wind power at time t on the predicted day, and are numbered 1-10, corresponding to the wind speed, wind direction, temperature, humidity, pressure, and the wind power at time t of the first, second, third, fourth, and fifth historical days on the predicted day. The Pearson correlation coefficients between the predicted daily wind power and the 10 initially selected features were calculated respectively. The PCC between the time series of the predicted daily wind power and the time series of each influencing feature is as follows: Figure 1 As shown.
[0079] like Figure 1 As shown, by comparing |PCC(x,y)| with a preset threshold, the correlation strength between the predicted daily wind power and its influence characteristics can be determined: assuming the preset threshold is 0.2, when |PCC(x,y)|∈[0,0.2], it represents extremely weak correlation or no correlation; when |PCC(x,y)|∈(0.2,0.4], it represents weak correlation; when |PCC(x,y)|∈(0.4,0.6], it represents moderate correlation; when |PCC(x,y)|∈(0.6,0.8], it represents strong correlation; and when |PCC(x,y)|∈[0.8,1], it represents extremely strong correlation. When PCC(x,y) is positive... When PCC(x,y) is positive, it represents a positive correlation; when PCC(x,y) is negative, it represents a negative correlation. Wind power fluctuates greatly, and many factors affect wind power output. To ensure the accuracy of wind power prediction, features with very weak correlation and no correlation need to be removed, and the remaining features are used to predict the wind power output on the day to be predicted. These features include wind speed, wind direction, temperature, pressure, historical wind power output on the first day, and historical wind power output on the second day, totaling six dimensions. Among these, temperature and pressure have a negative correlation with the wind power output on the day to be predicted, meaning that when other factors remain constant, wind power decreases when temperature and pressure increase. The other features have a positive correlation with the wind power output on the day to be predicted.
[0080] Based on the 6-dimensional features identified through data analysis, the expected corresponding features in the dataset are normalized, resulting in 5 normalized feature data and 2 trigonometric function-based feature data. Following the order of wind speed, wind direction sin value, wind direction cos value, temperature, pressure, historical wind power on the first day, historical wind power on the second day, and wind power on the day to be predicted, the normalized features are sequentially integrated with the wind power data for the day to be predicted into a single row containing 24×8 data points. Data from 365 days of the year is used to construct 363 original data samples. 250 original data samples are randomly selected to form the original dataset, and the remaining 113 original data samples are used as the test set. The sample format of the original dataset is as follows: Figure 2 As shown.
[0081] It should be noted that the preset threshold and the prediction time t for the day to be predicted can be set according to the actual prediction requirements.
[0082] Step S102: Remove the daily wind power to be predicted from the original dataset to obtain the clustered dataset and perform k-means clustering to obtain the original dataset with cluster labels. Generate an adversarial network based on robust auxiliary classification to generate a large number of labeled scenes.
[0083] The original dataset contains various features, including numerical weather prediction meteorological characteristics and historical daily wind power, with significant differences between different scenarios. To improve the accuracy of wind power prediction, the features affecting the wind power of the day to be predicted are clustered to obtain multi-cluster labels. Each scenario in the original dataset is then assigned a label. This facilitates targeted selection of similar scenarios within the corresponding cluster label-generated scenarios based on NWP (Numerical Weather Prediction) meteorological characteristics and historical daily wind power during prediction, thereby improving the accuracy of wind power prediction.
[0084] K-means clustering algorithm has the advantages of simple processing of large datasets and fast computation speed. It can achieve obvious feature differences between clusters and better clustering results. K-means clustering is to cluster n data objects into k classes (n and k are positive integers) to maximize the similarity of data objects in each class. This process is repeated until the division is completed. The algorithm steps are as follows: 1) Determine the number of clusters k and arbitrarily select k initial cluster centers from the data; 2) Calculate the distance from all data points to the cluster centers and classify the data into their respective classes according to the principle of minimum distance; 3) Based on the characteristics of each class, iteratively update the center value of each class using the mean method until the iteration ends.
[0085] To address the impact of NWP (Non-Wave Power Prediction) errors on wind power forecasting and the influence of limited training sample data on forecasting accuracy, this invention employs an improved robust auxiliary classifier generative adversarial network (RAC-GAN) to generate a massive amount of labeled wind power and its influencing features. In complex scenarios where NWP errors exist between the NWP and actual weather conditions, i.e., in noisy scenarios, to address the issues of unstable features affecting wind power, high extraction difficulty, and low wind power forecasting accuracy, and to meet the requirement of generating multi-label, robust scenarios under the influence of NWP errors, the following approach is proposed: Figure 3 The RAC-GAN model shown.
[0086] Auxiliary classifier generative adversarial networks (AC-GANs) add random noise signal labels and multi-classification functionality to the generative adversarial network (GAN) architecture, enabling them to generate samples of a specified type based on the labels. In the AC-GAN generator G, a random noise signal z and the corresponding label c of the generated sample are added. The generator specifically generates samples X(fake) = G(c,z) of the corresponding class. The discriminator D outputs the sample X derived from the real sample X(real), the probability P(S|X) of generating the sample X(fake), and the probability P(C|X) of belonging to different classes.
[0087] [P(S|X), P(C|X)] = D(X) (2)
[0088] In equation (2), S represents the sample source, X represents the output sample; C = c, c ∈ {1, 2, ..., n}, and n represents the number of sample classes. In AC-GAN, the objective function of G is to maximize L C -L S The objective function of D is to maximize L. C +L S L S With L C Defined as:
[0089]
[0090] In equation (3), L S The correct source loss function can determine the correctness of the data source; L C The loss function is the correct class loss function, which can determine the correctness of the output class; E is the expected value. Through the internal game between the generator and the discriminator, the generator is alternately optimized during the iterative process, ultimately improving the scene generation capability of the generator.
[0091] To meet the need for generating multi-label wind power scenarios, a data encoder is first introduced into the generator input of the AC-GAN. Instead of directly using random noise signals, the model pre-learns shallow raw data features based on real wind power and its impact characteristics, resulting in random noise input. This noise is then fed into the generator, enabling it to generate a massive amount of data that conforms to the probability distribution characteristics of real samples. The generated samples and original samples are then input into a discriminator to determine the quality of the generated samples, thus expanding the training sample data. During the RAC-GAN model iteration process, game-theoretic optimization is performed to reduce the impact of noise, ultimately achieving robust multi-label scenario generation under noise interference.
[0092] Step S103: Based on the known historical daily wind power and numerical weather forecast meteorological characteristics, determine the cluster label category of the day to be predicted, and based on the cluster label category, select multiple scenarios with high similarity to the wind power of the day to be predicted from a large number of labeled scenarios to form a similar scenario set.
[0093] The generated samples contain seven features affecting the wind power of the day to be predicted, along with the wind power of that day. Five meteorological factors can be obtained through NWP (Non-Warranty Predictors), and the historical wind power of the first and second days is known; therefore, all seven features affecting the wind power of the day to be predicted are known. To determine which cluster label the influencing features belong to, based on the weighted PCC (Plan-Do-Check-Act) theory, multiple scenarios with high similarity to the wind power of the day to be predicted are selected from the generated samples of that cluster label, forming a similar scenario set.
[0094] Step S104: Based on the average, upper limit and lower limit of the wind power on the day to be predicted in the similar scenario set, obtain the point prediction and interval prediction results of the wind power at multiple set times on the day to be predicted.
[0095] In practical applications, the PCC (Potential Constituent Control) of different influencing features and the wind power on the predicted day varies, thus the degree to which each influencing feature affects the wind power on the predicted day differs. Using the PCC of the wind power on the predicted day and each influencing feature as weighting coefficients, the weighted PCC of all influencing features of the wind power on the predicted day and the corresponding features of the j-th generated sample is set as the PCC. j The expression is shown in equation (4):
[0096]
[0097] In equation (4), PCC(x, y) i ) represents the PCC (predicted daily wind power output) and characteristic i; PCC(y) i y ji () represents the PCC between known feature i and feature i of the j-th generated sample; PCC j is the weighted PCC between the j-th generated sample and the features of the day to be predicted.
[0098] Given the various influencing characteristics of wind power on a given day to be predicted, firstly, the cluster labels corresponding to these characteristics are analyzed based on the k-means clustering algorithm to find the generated samples corresponding to the cluster labels; then, according to equation (4), according to PCC... j In descending order, N samples with high inter-day similarity to the historical data of the day to be predicted are selected from the cluster label generation samples to form a similar scene set. The wind power scene of the day to be predicted in the similar scene set is extracted and defined as set Z. Finally, based on the average value, upper limit value and lower limit value of each time period in Z, the point prediction and interval prediction results of wind power at multiple set times of the day to be predicted are obtained, as shown in Equation (5):
[0099]
[0100] In equation (5), P wtmax，t P wtmin，t P represents the upper and lower limits of the predicted wind power range at time t; wtmean，t This represents the point prediction result of wind power at time t.
[0101] In practical applications, in order to determine the required sample for wind power prediction, the prediction interval coverage proportion (PICP) and prediction interval normalized average width (PINAW) are used to analyze the interval prediction effect under different sample numbers, and the mean absolute percentage error (MAPE) is used to analyze the point prediction effect under different sample numbers.
[0102] To verify the prediction effect of the prediction method in this embodiment of the invention, experimental analysis was conducted using one year's measured wind farm data from a region in Northeast China.
[0103] The number of samples selected was set to 1%, 2%, 3%, 4%, and 5% of the number of generated samples included in each cluster label, respectively. The prediction indicators under different numbers of selected samples are shown in Table 1.
[0104] Analyzing Table 1, the optimal screening ratio for the indicator cannot be directly selected. Therefore, it is stipulated that the indicator can receive 5 points when it is optimal under the five screening ratios, and 1 point when it is worst. According to this rule, the scores under different screening ratios are shown in the rightmost column of Table 1. The prediction effect is optimal when the number of screened samples is 4% of the number of samples generated for the corresponding cluster label.
[0105] Table 1
[0106]
[0107]
[0108] Since the uncertainty of wind power is inherently non-Gaussian, traditional Gaussian processes are not suitable for wind power prediction. Therefore, a warped Gaussian process (WGP), which has shown better performance in predicting wind power over time intervals, was used for comparative experiments. The experiments were conducted in Matlab R2015b with confidence levels set to 90%, 95%, and 99%. By randomly selecting one week from each of the four seasons as the prediction experiment, the prediction indices of the different prediction methods are shown in Table 2 below.
[0109] Table 2
[0110] Prediction methods MAPE / % PICP / % PINAW / % WGP (90% confidence level) 17.8 85.2 18.2 WGP (95% confidence level) 17.8 89.2 22.1 WGP (99% confidence level) 17.8 90.1 25.0 Method of this invention 13.4 96.3 15.4
[0111] Analysis of Table 2 shows that the prediction indicators of the method in this embodiment are all superior to those of the WGP prediction method, proving that the method in this embodiment has better interval prediction performance. Wind power is significantly affected by the seasons; therefore, the wind power prediction performance of the method in this embodiment and the WGP method at different confidence levels is analyzed in each season. A day is randomly selected in each season for prediction, and the prediction results of different methods are shown below. Figure 6 As shown. By Figure 6 It can be seen that, in the four days randomly selected from each season, the prediction method of this embodiment of the invention has the lowest MAPE and PINAW, and the highest PICP. This is because a massive number of generated scenarios are first obtained through scenario generation, and then similar scenarios are filtered by sub-labels. The obtained similar scenarios have a high similarity to the real scenarios, so the prediction error percentage is low, the prediction interval width is low, and the prediction interval coverage is high. In summary, in the study of wind power day-ahead prediction, whether the experiment selects one week from each of the four seasons or randomly selects one day from each of the four seasons, the method of this embodiment of the invention has certain advantages.
[0112] The features used in wind power prediction include meteorological features obtained from NWP (Non-Warranty Prediction). There is an error between the NWP meteorological features and the actual weather conditions, and the magnitude of this error directly affects the quality of selecting similar scenarios, thus indirectly affecting the prediction effect. To analyze the prediction effect of the method in this embodiment under different NWP errors, noise was added during RAC-GAN scenario generation to simulate NWP errors, and signal-to-noise ratios of 20dB, 25dB, 30dB, 35dB, and 40dB were set to simulate NWP errors. The interval prediction indices obtained by the new wind power interval prediction method based on label-based screening are shown in Table 3 below.
[0113] Table 3
[0114] Signal-to-noise ratio (dB) MAPE / % PICP / % PINAW / % 20 16.7 92.6 19.2 25 16.4 94.3 18.1 30 16.0 95.4 17.4 35 15.2 96.0 16.7 40 14.3 96.6 16.1
[0115] Analyzing Table 3, as the signal-to-noise ratio decreases, the noise becomes stronger, meaning the NWP error is larger, which is increasingly detrimental to the accuracy and stability of the prediction results. Specifically, the prediction effect is worst at 20dB, but the prediction index is still better than the WGP prediction method at this point. This proves that the prediction method of the present invention can still guarantee good interval prediction results even when there is a large NWP error, demonstrating the practicality of the new wind power interval prediction method based on RAC-GAN scenario generation in the present invention.
[0116] The wind power prediction method for wind farms provided in this invention constructs an original dataset containing the wind power of the day to be predicted, meteorological characteristics from numerical weather prediction, and historical wind power using the Pearson correlation coefficient. Based on a clustering algorithm, a labeled original dataset is constructed. A scene generation model based on RAC-GAN is designed to generate a massive number of labeled samples. The cluster label category for the day to be predicted is determined based on the known characteristics of the wind power. Scenes with high feature similarity to this cluster label are selected, thereby obtaining point and interval prediction results for wind power, improving the accuracy of point and interval prediction. Compared with the traditional WGP interval prediction method, the prediction method of this invention is more accurate and effective. Furthermore, considering the difference between NWP and actual meteorological conditions, noise is incorporated to simulate NWP error when constructing the RAC-GAN-based scene generation model, and prediction results are obtained under this condition. Simulation verification shows that NWP error slightly reduces the accuracy and reliability of the prediction results, but the new method considering NWP error still performs better than the WGP method, verifying that the method of this invention has a stronger ability to withstand NWP error and the prediction results are less affected by external factors.
[0117] See Figure 7 This invention also provides a day-ahead wind power prediction system for wind farms, comprising:
[0118] Module 701 can be used to construct an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted, based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power.
[0119] The clustering generation module 702 can be used to remove the predicted daily wind power from the original dataset, obtain the clustered dataset, perform k-means clustering, obtain the original dataset with cluster labels, and generate a massive number of labeled scenes by using robust auxiliary classification to generate an adversarial network.
[0120] The filtering module 703 can be used to determine the cluster label category of the day to be predicted based on the known historical daily wind power and the meteorological characteristics of numerical weather forecasts, and to filter out multiple scenarios with high similarity to the wind power of the day to be predicted from a large number of labeled scenarios based on the cluster label category, and form a similar scenario set.
[0121] The prediction module 704 can be used to obtain point prediction and interval prediction results of wind power at multiple set times on the day to be predicted based on the average, upper limit and lower limit of wind power on the day to be predicted in a similar scenario set.
[0122] The construction module 701 further includes:
[0123] The first selection unit can be used to select wind speed, wind direction, temperature, humidity and pressure as the first preliminary features for the wind power of the day to be predicted;
[0124] The second selection unit can be used to select the wind power of historical days with an absolute value of Pearson correlation coefficient greater than a preset threshold as the second preliminary feature of wind power of the day to be predicted, based on the correlation between the wind power of each historical day and the wind power of the day to be predicted.
[0125] The calculation and determination unit can be used to calculate the Pearson correlation coefficient between the predicted daily wind power and the first and second preliminary features, and compare the absolute value of the Pearson correlation coefficient with the preset threshold to determine the correlation between the predicted daily wind power and the first and second preliminary features.
[0126] The cell removal function can be used to remove weakly correlated or uncorrelated initial features based on the cell determination results obtained from calculations.
[0127] The normalization unit can be used to take the initial selected features remaining after the removal unit is removed as the influence features of the predicted daily wind power, and normalize the features of the dataset that correspond to the influence features.
[0128] The integration unit can be used to integrate the normalized features with the wind power of the day to be predicted in a pre-set arrangement order to form the original data of the day to be predicted.
[0129] The dataset unit can be used to construct a sample of raw data for the whole year based on the raw data of each day to be predicted, and to randomly select some raw data from the sample of raw data for the whole year to construct the raw dataset.
[0130] The clustering generation module 702 further includes:
[0131] Clustered dataset units can be used to remove the predicted daily wind power from the original dataset, obtain the normalized features corresponding to the influencing features, and form a clustered dataset.
[0132] Clustering units can be used to perform k-means clustering on data in a clustered dataset to obtain multi-cluster labels;
[0133] Assigning units can be used to assign labels to each scene in the original dataset;
[0134] The learning unit can be used to introduce a data encoder at the generator input of the auxiliary classification generative adversarial network, and to pre-learn shallow raw data features based on real wind power and its impact characteristics data to obtain random noise input;
[0135] The generator unit can be used to generate massive amounts of sample data that satisfy the probability distribution characteristics of real samples based on random noise.
[0136] The discriminator unit can be used to determine the quality of the generated samples based on the sample data generated by the generator unit and the original sample data, thereby expanding the training sample data.
[0137] The optimized generation unit can be used in the iterative process of robust auxiliary classification generative adversarial network to carry out game optimization in the direction of reducing the impact of noise, and finally generate a large number of labeled wind power and its impact characteristics under noise interference.
[0138] The screening module 703 further includes:
[0139] The search unit can be used to find the corresponding generated sample in a massive labeled scenario based on the cluster label of the predicted date;
[0140] The setting unit can be used to set the weighted PCC of all influence features of the wind power to be predicted on the day of wind power and the influence features as the weighting coefficients, and to set the PCC of the corresponding features between the j-th generated sample and the Pearson correlation coefficient between the wind power to be predicted on the day of wind power. j ;
[0141] Similar scene set units can be used based on PCC j From largest to smallest, select multiple samples with high similarity to historical days that are highly correlated with the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set.
[0142] The prediction module 704 further includes:
[0143] The extraction unit can be used to extract the predicted daily wind power scenarios from a set of similar scenarios and define them as a set Z;
[0144] The prediction unit can be used to obtain point prediction and interval prediction results of wind power at multiple set times on the day to be predicted, based on the average value of each time period and the upper and lower limits in set Z.
[0145] Furthermore, the wind farm day-ahead wind power prediction system provided in this embodiment of the invention also includes:
[0146] The first screening module can be used to analyze the point prediction results of wind power under different screening sample numbers using the mean absolute percentage error.
[0147] The second screening module can be used to analyze the interval prediction results of wind power under different screening sample numbers by using the prediction interval coverage and the prediction interval normalized average width.
[0148] It should be noted that other corresponding descriptions of the functional modules involved in the wind farm day-ahead wind power prediction system provided in this embodiment of the invention can be found in the following references. Figure 4 and 5 The corresponding description of the method shown will not be repeated here.
[0149] Based on the above, Figure 4 and Figure 5 Accordingly, this embodiment of the invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the following steps: Based on the correlation between the wind power of the day to be predicted, numerical weather forecast meteorological characteristics, and historical daily wind power, construct an original dataset containing numerical weather forecast meteorological characteristics and the wind power of the day to be predicted; remove the wind power of the day to be predicted from the original dataset to obtain a clustered dataset and perform k-means clustering to obtain an original dataset with cluster labels; generate a massive number of labeled scenes using a robust auxiliary classification generative adversarial network; determine the cluster label category of the day to be predicted based on the known historical daily wind power and numerical weather forecast meteorological characteristics; and, based on the cluster label category, select multiple scenes with high similarity to the wind power of the day to be predicted from the massive number of labeled scenes to form a similar scene set; based on the average, upper limit, and lower limit of the wind power of the day to be predicted in the similar scene set, obtain point prediction and interval prediction results of the wind power at multiple set times on the day to be predicted.
[0150] Based on the above, Figure 4 , 5 The method shown and as Figure 7 The embodiment of the system shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 8As shown, the computer device includes a processor 81, a memory 82, and a computer program stored in the memory 82 and executable on the processor. Both the memory 82 and the processor 81 are mounted on a bus 83. When the processor 81 executes the program, it performs the following steps: Based on the correlation between the wind power of the day to be predicted, numerical weather forecast meteorological characteristics, and historical daily wind power, it constructs an original dataset containing the numerical weather forecast meteorological characteristics and the wind power of the day to be predicted; it removes the wind power of the day to be predicted from the original dataset, obtains a clustered dataset, performs k-means clustering, obtains an original dataset with cluster labels, and generates a massive number of labeled scenes using a robust auxiliary classification generative adversarial network; based on the known historical daily wind power and numerical weather forecast meteorological characteristics, it determines the cluster label category of the day to be predicted, and based on the cluster label category, it selects multiple scenes with high similarity to the wind power of the day to be predicted from the massive number of labeled scenes to form a similar scene set; based on the average, upper limit, and lower limit of the wind power of the day to be predicted in the similar scene set, it obtains point prediction and interval prediction results of the wind power at multiple set times on the day to be predicted.
[0151] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.
[0152] In practical applications, each functional module and unit involved in the embodiments of the present invention can be implemented by a computer program running on computer hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the flow of the embodiments of the above methods. The hardware refers to a server, desktop computer, laptop computer, etc., containing one or more processors and storage media; the storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.; the computer program is implemented in computer languages not limited to C, C++, etc.
[0153] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting day-ahead wind power in a wind farm, characterized in that, include: Based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power, an original dataset containing the meteorological characteristics of numerical weather forecasts and the wind power of the day to be predicted is constructed. Remove the predicted daily wind power from the original dataset to obtain a clustered dataset and perform k-means clustering to obtain the original dataset with cluster labels. Generate an adversarial network based on robust auxiliary classification to generate a large number of labeled scenarios. The steps of generating a massive number of labeled scenes using a robust auxiliary classification generative adversarial network specifically include: A data encoder is introduced at the generator input of the auxiliary classification generative adversarial network, and shallow raw data features are pre-learned based on real wind power and its impact characteristics to obtain random noise input. The generator generates a massive amount of sample data that satisfies the probability distribution characteristics of real samples based on the random noise. The discriminator judges the quality of the generated samples based on the generated sample data and the original sample data, thereby expanding the training sample data; The robust auxiliary classification generative adversarial network iterates through a game-theoretic optimization process to reduce the impact of noise, and finally generates a massive number of labeled wind power and its impact characteristics under noise interference. Based on known historical daily wind power and numerical weather forecast meteorological characteristics, the cluster label category of the day to be predicted is determined. Then, based on the cluster label category, multiple scenarios with high similarity to the wind power of the day to be predicted are selected from the massive number of labeled scenarios to form a similar scenario set. The specific steps of selecting multiple scenarios with high similarity to the wind power of the day to be predicted based on the cluster label category to form a similar scenario set include: Based on the cluster label of the predicted date, the corresponding generated sample is found in the massive labeled scenario; Using the Pearson correlation coefficient between the predicted daily wind power and the influencing features as the weighting coefficient, the weighted PCC of all influencing features of the predicted daily wind power and the corresponding features of the j-th generated sample is set as PCC. j The expression is: Where PCC(x,y) i ) represents the PCC between the predicted daily wind power and the influencing characteristic i; PCC(y) i ,y ji () represents the PCC between known influence feature i and influence feature i of the j-th generated sample; PCC j The weighted PCC between the j-th generated sample and the features of the day to be predicted; According to the PCC j From largest to smallest, select multiple samples with high historical daytime similarity that are highly correlated with the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set; Based on the average, upper limit, and lower limit of the wind power on the day to be predicted in the similar scenario set, point prediction and interval prediction results of wind power at multiple set times on the day to be predicted are obtained respectively.
2. The wind farm day-ahead wind power prediction method according to claim 1, characterized in that, The specific steps of basing the prediction of daily wind power output, numerical weather forecast meteorological characteristics, and historical daily wind power output on the correlation include: Wind speed, wind direction, temperature, humidity, and pressure were selected as the first preliminary features for predicting daily wind power. Based on the correlation between the wind power of each historical day and the wind power of the day to be predicted, the wind power of historical days with an absolute value of Pearson correlation coefficient greater than a preset threshold is selected as the second preliminary feature of the wind power of the day to be predicted. Calculate the Pearson correlation coefficient between the predicted daily wind power and the first and second preliminary features; By comparing the absolute value of the Pearson correlation coefficient with a preset threshold, the correlation between the predicted daily wind power and the first and second preliminary features is determined.
3. The wind farm day-ahead wind power prediction method according to claim 2, characterized in that, The formula for calculating the Pearson correlation coefficient is as follows: Where x is the daily wind power to be predicted, y i The influence characteristics of the predicted daily wind power output; x k y ik x and y are respectively i The k-th data point; x and y are respectively i The average value of the data; k and n are positive integers.
4. The wind farm day-ahead wind power prediction method according to claim 3, characterized in that, The steps for constructing the original dataset containing the numerical weather forecast meteorological features and the daily wind power to be predicted specifically include: Based on the correlation determination results between the predicted daily wind power and the first and second preliminary features, the preliminary features with very weak correlation or no correlation are removed. The remaining initially selected features are used as the influence features for predicting the wind power of the day to be predicted, and the features in the dataset that are expected to correspond to the influence features are normalized. According to the pre-set arrangement order, the normalized features are integrated with the wind power of the day to be predicted to form the original data of the day to be predicted. Based on the raw data for each day to be predicted, construct a sample of raw data for the whole year, and randomly select a number of raw data from the sample of raw data for the whole year to construct the raw dataset.
5. The wind farm day-ahead wind power prediction method according to claim 4, characterized in that, The steps of removing the predicted daily wind power from the original dataset, obtaining a clustered dataset, and performing k-means clustering to obtain the original dataset with cluster labels specifically include: By removing the predicted daily wind power from the original dataset, the normalized features corresponding to the influencing features are obtained, forming a clustered dataset. The data in the clustered dataset are subjected to k-means clustering to obtain multi-cluster labels; Assign the label to each scene in the original dataset.
6. The wind farm day-ahead wind power prediction method according to claim 5, characterized in that, The k-means clustering method clusters n data objects into k classes, maximizing the similarity of data objects in each class. This process is repeated until the division is complete. The steps are as follows: 1) Determine the number of clusters k and arbitrarily select k initial cluster centers from the data; 2) Calculate the distance from all data points to the cluster centers and classify the data into their respective categories according to the principle of minimum distance; 3) Based on the characteristics of each class, iteratively update the centers of each class using the mean method until the iteration ends.
7. The wind farm day-ahead wind power prediction method according to claim 6, characterized in that, The step of obtaining point prediction and interval prediction results of wind power at multiple set times on the day to be predicted based on the average, upper limit and lower limit of wind power on the day to be predicted in the similar scenario set specifically includes: The wind power scenarios for the day to be predicted are extracted from the set of similar scenarios and defined as set Z; based on the average value and upper and lower limits of each time period in set Z, the point prediction and interval prediction results of wind power at multiple set times on the day to be predicted are obtained, as shown in the following formula: P wtmax,t =a(1,t) a=max[Z] P wtmin,t =b(1,t) b=min[Z] t=(1,2,..,24) P wtmean,t =c(1,t) c=mean[Z] Among them, P wtmax,t P wtmin,t P represents the upper and lower limits of the wind power range prediction at time t; wtmean,t This represents the point prediction result of wind power at time t.
8. The wind farm day-ahead wind power prediction method according to claim 7, characterized in that, The method further includes: using the mean absolute percentage error to analyze the point prediction results of the wind power under different numbers of selected samples, and using the prediction interval coverage and the normalized average width of the prediction interval to analyze the interval prediction results of the wind power under different numbers of selected samples.
9. A wind farm day-ahead wind power prediction system, characterized in that, include: The construction module is used to construct an original dataset containing the numerical weather forecast meteorological characteristics and the wind power of the day to be predicted, based on the correlation between the wind power of the day to be predicted, the meteorological characteristics of numerical weather forecasts, and the historical wind power. The clustering generation module is used to remove the predicted daily wind power from the original dataset, obtain a clustered dataset, perform k-means clustering, obtain an original dataset with cluster labels, and generate an adversarial network based on robust auxiliary classification to generate a large number of labeled scenes. The filtering module is used to determine the cluster label category of the day to be predicted based on the known historical daily wind power and the meteorological characteristics of numerical weather forecasts, and to filter out multiple scenarios with high similarity to the wind power of the day to be predicted from the massive number of labeled scenarios based on the cluster label category, forming a similar scenario set; The prediction module is used to obtain point prediction and interval prediction results of wind power at multiple set times on the day to be predicted based on the average value, upper limit value and lower limit value of wind power on the day to be predicted in the similar scenario set. The clustering generation module includes: Clustering dataset unit is used to remove the predicted daily wind power from the original dataset to obtain the normalized features corresponding to the influencing features, thus forming a clustering dataset; Clustering unit, used to perform k-means clustering on the data in the clustered dataset to obtain multi-cluster labels; A labeling unit is used to assign the label to each scene in the original dataset. The learning unit is used to introduce a data encoder at the generator input of the auxiliary classification generative adversarial network, and to pre-learn shallow raw data features based on real wind power and its impact characteristics to obtain random noise input. The generator unit is used to generate massive amounts of sample data that satisfy the probability distribution characteristics of real samples based on the random noise. The discriminator unit is used to determine the quality of the generated samples based on the sample data generated by the generator unit and the original sample data, thereby expanding the training sample data. The optimized generation unit is used to assist the classification generative adversarial network in the iterative process, and to carry out game optimization in the direction of reducing the impact of noise, and finally generate a large number of labeled wind power and its impact characteristics under noise interference. The filtering module includes: The search unit is used to find the corresponding generated sample in the massive labeled scenario based on the cluster label of the predicted day. The setting unit is used to set the weighted PCC of all influence features of the wind power to be predicted on the day of wind power and the influence features as the weighting coefficients, and the corresponding features between the j-th generated sample and the Pearson correlation coefficient between the wind power to be predicted on the day of wind power. j ; Similar scene set unit, used according to the PCC j From largest to smallest, select multiple samples with high inter-day similarity to the historical data of the day to be predicted from the generated samples corresponding to the cluster label, and form a similar scene set.
10. The wind farm day-ahead wind power prediction system according to claim 9, characterized in that, The building module includes: The first selection unit is used to select wind speed, wind direction, temperature, humidity and pressure as the first preliminary features for the wind power of the day to be predicted; The second selection unit is used to select the wind power of historical days with an absolute value of Pearson correlation coefficient greater than a preset threshold as the second preliminary feature of the wind power of the day to be predicted, based on the correlation between the wind power of each historical day and the wind power of the day to be predicted. The calculation and determination unit is used to calculate the Pearson correlation coefficient between the predicted daily wind power and the first preliminary feature and the second preliminary feature, and compare the absolute value of the Pearson correlation coefficient with the preset threshold to determine the correlation between the predicted daily wind power and the first preliminary feature and the second preliminary feature. The removal unit is used to remove initially selected features that are extremely weakly correlated or have no correlation, based on the determination result of the calculation unit. The normalization unit is used to take the initial selected features remaining after the removal unit as the influence features of the predicted wind power on the day to be predicted, and to normalize the features of the dataset that correspond to the influence features. An integration unit is used to integrate the normalized features with the wind power of the day to be predicted in a pre-set arrangement order to form the original data of the day to be predicted. The dataset unit is used to construct a sample of raw data for the whole year based on the raw data of each day to be predicted, and to randomly select a number of raw data from the sample of raw data for the whole year to construct the raw dataset.
11. The wind farm day-ahead wind power prediction system according to claim 10, characterized in that, The prediction module includes: The extraction unit is used to extract the predicted daily wind power scene from the set of similar scenes and define it as set Z; The prediction unit is used to obtain point prediction and interval prediction results of wind power at multiple set times on the day to be predicted based on the average value and upper and lower limits of each time period in the set Z.
12. The wind farm day-ahead wind power prediction system according to claim 11, characterized in that, The system also includes: The first screening module is used to analyze the point prediction results of the wind power under different screening sample numbers using the mean absolute percentage error. The second screening module is used to analyze the interval prediction results of the wind power under different screening sample numbers by using the prediction interval coverage and the prediction interval normalized average width.
13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.
14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.