Wind power ultra-short-term power prediction method based on machine learning
Through the dual SVM model series architecture, the cloud mask features of the wind farm are extracted and the nonlinear relationship between cloud features and power output is learned, which solves the problems of insufficient accuracy and poor adaptability of traditional wind power power prediction methods, and achieves higher accuracy and real-time prediction effects.
Patent Information
- Application Number
- CN202411831129.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-05-13
Smart Images
Figure CN119994849A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind turbines, and in particular to a method for predicting wind power ultra-short-term power based on machine learning. Background Art
[0002] With the rapid development of the wind power industry, accurate wind power forecasting is of great significance to the safe and stable operation of the power grid and the economic benefits of wind farms. As a key meteorological factor affecting the power generation efficiency of wind farms, the distribution and changes of clouds directly affect the power output of wind farms. All-sky imagers can capture cloud images above wind farms in real time, providing an important data basis for cloud analysis and power forecasting. At present, traditional wind power forecasting methods mainly rely on numerical weather forecast (NWP) data, which have limitations such as low temporal and spatial resolution and slow update frequency. The forecasting method based on meteorological stations cannot accurately reflect the cloud changes in the local area of the wind farm, resulting in insufficient forecasting accuracy. The statistical forecasting method that relies solely on historical power data ignores the influence of important meteorological factors such as clouds, and the forecasting effect is poor when the weather changes drastically. In addition, traditional cloud detection algorithms often use fixed threshold methods, which are difficult to adapt to different weather conditions and lighting environments, and are prone to misjudgment, affecting the accuracy of subsequent power forecasting. Based on this, in response to the above problems, we designed a wind power ultra-short-term power forecasting method based on machine learning. Summary of the invention
[0003] The purpose of the present invention is to provide a wind power ultra-short-term power prediction method based on machine learning, which adopts an innovative architecture of dual SVM models in series. The cloud mask features extracted by the first SVM classifier directly reflect the cloud distribution status above the wind farm, providing more effective input features for power prediction. The nonlinear relationship between cloud features and power output is learned by the second SVM model, thereby achieving more accurate prediction. The method not only has strong real-time performance, good adaptability and high prediction accuracy, but also can provide more reliable decision support for the operation scheduling and power grid regulation of wind farms.
[0004] The present invention is achieved through the following technical solutions:
[0005] A wind power ultra-short-term power prediction method based on machine learning, the method comprising the following steps:
[0006] Obtain all-sky image data and power data of wind farms;
[0007] Perform image processing on all-sky image data of wind farms;
[0008] Extracting features from the wind farm full sky image data after image processing, and normalizing the features of the wind farm full sky image data;
[0009] Constructing a first SVM classifier model, taking the features of the wind farm full sky image data as input of the first SVM classifier model, outputting the classification results of the wind farm full sky image data, and screening and outputting the cloud mask image data;
[0010] The power characteristics of the wind farm are calculated based on the cloud mask image data, a second SVM classifier model is constructed, the power characteristics of the wind farm are used as input of the second SVM classifier model, and the predicted power value of the wind farm is output.
[0011] Optionally, the image processing of the full sky image data of the wind farm is specifically: based on synchronously acquiring the full sky image data of the wind farm based on multiple adjacent devices, calculating the Euclidean distance of the RGB channel histogram between the devices, identifying the device whose Euclidean distance exceeds the set value as an abnormal device, and performing histogram equalization processing on the detected abnormal device image, and completing the image processing of the full sky image data of the wind farm through multiple iterative optimization and correction.
[0012] Optionally, the features of the wind farm full sky image data include: RGB three-channel value features, red-to-blue ratio features, Gaussian spatial filtering features based on a set size window, and brightness channel standard deviation features within a set size window.
[0013] Optionally, the training process of the first SVM classifier model is:
[0014] Obtain historical all-sky image data of the wind farm and the corresponding manually labeled cloud / sky pixel labels, and perform image processing and feature extraction;
[0015] The features and corresponding labels of the historical all-sky image data of the wind farm are divided into a training set and a test set;
[0016] Construct the first SVM classifier model, define the kernel function, optimization objectives and constraints of the first SVM classifier model, set the parameter search range, divide the training set into 5 parts, and cycle each training subset 5 times, with 1 part for validation and the other 4 parts for training;
[0017] Calculate the verification accuracy and take the average accuracy of 5 times as the performance indicator of this group of parameters. Select the parameter combination with the highest average accuracy to train on the training set to obtain the decision function. Compare the prediction results in the decision function with the true labels of the test set and output the first SVM classifier model that has completed the training.
[0018] Optionally, the kernel function of the first SVM classifier model is calculated as follows:
[0019] K(x i ,x j )=exp(-γ‖x i -xj ‖ 2 )
[0020] The optimization objective and constraint conditions of the first SVM classifier model are calculated as follows:
[0021]
[0022]
[0023] 0≤α i ≤C,i=1,...,n
[0024] Among them, K(x,x') is the RBF kernel function, α i , α j are the Lagrange multipliers of the i-th and j-th sample points respectively, α is the Lagrange multiplier set, x i is the feature vector of the i-th training sample, γ is the kernel function parameter, y i is the label of the i-th sample, n is the total number of samples, and C is the penalty parameter.
[0025] Optionally, the decision function is calculated as follows:
[0026]
[0027] Where b is the bias.
[0028] Optionally, the training process of the second SVM classifier model is:
[0029] Perform cloud detection on the historical full-sky image data of the wind farm through the first SVM classifier model to obtain historical cloud mask image data. Based on the historical cloud mask image data, calculate the power characteristics of the wind farm: cloud coverage, cloud motion characteristics, cloud type distribution, and collect historical power values at the corresponding time.
[0030] After preprocessing, the power characteristics and power values of the wind farm are divided into a training data set and a test data set in chronological order;
[0031] Set the parameter search range, divide the training data set into 5 parts evenly, and cycle through each training data subset 5 times, with 1 part for validation and the other 4 for training;
[0032] The verification accuracy is calculated, and the average accuracy of 5 times is taken as the performance indicator of this group of parameters. The parameter combination with the highest average accuracy is selected for training on the training data set to obtain the decision function. The prediction results in the decision function are compared with the true labels of the test data set, and the second SVM classifier model that has completed the training is output.
[0033] Optionally, the loss function of the second SVM classifier model adopts RMSE, and its calculation formula is:
[0034]
[0035] Among them, y′ i is the actual power value, is the predicted power value.
[0036] The technical solution of the present invention has at least the following advantages and beneficial effects:
[0037] The present invention adopts an innovative architecture of dual SVM models in series. The cloud mask features extracted by the first SVM classifier directly reflect the cloud distribution above the wind farm, providing more effective input features for power prediction. The nonlinear relationship between cloud features and power output is learned by the second SVM model, thereby achieving more accurate prediction. It not only has strong real-time performance, good adaptability and high prediction accuracy, but also can provide more reliable decision support for the operation scheduling of wind farms and power grid regulation. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 A schematic flow chart of the method for wind power ultra-short-term power prediction based on machine learning provided by the present invention. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, what is described is only a part of the present invention, not all of it. Generally, the components of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0040] like Figure 1 As shown, the present invention provides one embodiment: a method for ultra-short-term wind power prediction based on machine learning, the method comprising the following steps:
[0041] Obtain all-sky image data and power data of wind farms;
[0042] Perform image processing on all-sky image data of wind farms;
[0043] Extracting features from the wind farm full sky image data after image processing, and normalizing the features of the wind farm full sky image data;
[0044] Constructing a first SVM classifier model, taking the features of the wind farm full sky image data as input of the first SVM classifier model, outputting the classification results of the wind farm full sky image data, and screening and outputting the cloud mask image data;
[0045] The power characteristics of the wind farm are calculated based on the cloud mask image data, a second SVM classifier model is constructed, the power characteristics of the wind farm are used as input of the second SVM classifier model, and the predicted power value of the wind farm is output.
[0046] In this embodiment, the image processing of the full sky image data of the wind farm is specifically as follows: the full sky image data of the wind farm is synchronously acquired based on multiple adjacent devices, the Euclidean distance of the RGB channel histogram between the devices is calculated, the device whose Euclidean distance exceeds the set value is identified as an abnormal device, and the histogram equalization processing is performed on the detected abnormal device image, and the image processing of the full sky image data of the wind farm is completed through multiple iterative optimization and correction.
[0047] In implementation, the number of devices described in this embodiment is set to three, and the synchronous images of three adjacent devices are used for multi-source anomaly correction. Since the positions of the three devices are quite close, it is reasonable to assume that their color expressions have statistically similar ranges in the RGB channels. That is, the cloud / sky pixels in the digitized color channels of the three different devices should have similar histograms. Therefore, if a device has an exposure problem or brightness abnormality in the RGB color space, we can correct it by equalizing the histogram of its RGB channel with the histograms of two normal devices. First, a cloud mask is generated for all images using an SVM classifier, and the histograms of the RGB channels of cloud pixels and sky pixels are calculated respectively. By calculating the Euclidean distance between the histogram vectors of the three devices, if the image histogram of a device is significantly different from that of the other two devices, we can identify it as an anomaly. Then, the histogram equalization is additionally applied to the output image of the abnormal device to adjust the RGB scale of its cloud and sky pixels. The corrected result can be used for the generation of the next round of cloud masks. In practice, this process is iterated three times to extract the cloud mask and balance the RGB histogram of the device image.
[0048] Specifically, the features of the wind farm full sky image data include: RGB three-channel value features, red-to-blue ratio features, Gaussian spatial filtering features based on a set size window, and brightness channel standard deviation features within a set size window.
[0049] In this embodiment, the training process of the first SVM classifier model is:
[0050] Obtain historical all-sky image data of the wind farm and the corresponding manually labeled cloud / sky pixel labels, and perform image processing and feature extraction;
[0051] The features and corresponding labels of the historical all-sky image data of the wind farm are divided into a training set and a test set;
[0052] Construct the first SVM classifier model, define the kernel function, optimization objectives and constraints of the first SVM classifier model, set the parameter search range, divide the training set into 5 parts, and cycle each training subset 5 times, with 1 part for validation and the other 4 parts for training;
[0053] Calculate the verification accuracy and take the average accuracy of 5 times as the performance indicator of this group of parameters. Select the parameter combination with the highest average accuracy to train on the training set to obtain the decision function. Compare the prediction results in the decision function with the true labels of the test set and output the first SVM classifier model that has completed the training.
[0054] The kernel function of the first SVM classifier model is calculated as follows:
[0055] K(x i ,x j )=exp(-γ‖x i -x j ‖ 2 )
[0056] The optimization objective and constraint conditions of the first SVM classifier model are calculated as follows:
[0057]
[0058] 0≤α i ≤C,i=1,...,n
[0059] Among them, K(x,x′) is the RBF kernel function, α i , α j are the Lagrange multipliers of the i-th and j-th sample points respectively, α is the Lagrange multiplier set, x i is the feature vector of the i-th training sample, γ is the kernel function parameter, y i is the label of the i-th sample, n is the total number of samples, and C is the penalty parameter.
[0060] The decision function is calculated as follows:
[0061]
[0062] Where b is the bias.
[0063] The performance of the supervised classifiers and multi-source correction algorithms described above are evaluated. Various test cases are selected from daily observations of different atmospheric conditions and cloud types, and the results are then compared with manually annotated images. In this report, two evaluation metrics are used to measure the error in cloud classification:
[0064]
[0065] Where ACcld and ACsky are the classification accuracy of cloud and sky pixels, respectively, Ncld,cld and Nsky,sky represent the pixel counts of correct cloud classification and sky classification, respectively, and Nsky,cld and Ncld,sky represent the total number of sky and cloud pixels that are misidentified by the detector, respectively. Compared with the manual classification mask, the SVM classifier-based pipeline can accurately detect clouds except for multi-layer clouds and very thin clouds near the sun position with an accuracy rate higher than 83.2%. .The image area close to the sun has a higher brightness and it is difficult to describe its features based on static texture information alone. Therefore, the classifier often mistakenly labels clouds in this area as sky pixels.
[0066] In this embodiment, the training process of the second SVM classifier model is:
[0067] Perform cloud detection on the historical full-sky image data of the wind farm through the first SVM classifier model to obtain historical cloud mask image data. Based on the historical cloud mask image data, calculate the power characteristics of the wind farm: cloud coverage, cloud motion characteristics, cloud type distribution, and collect historical power values at the corresponding time.
[0068] After preprocessing, the power characteristics and power values of the wind farm are divided into a training data set and a test data set in chronological order;
[0069] Set the parameter search range, divide the training data set into 5 parts evenly, and cycle through each training data subset 5 times, with 1 part for validation and the other 4 for training;
[0070] The verification accuracy is calculated, and the average accuracy of 5 times is taken as the performance indicator of this group of parameters. The parameter combination with the highest average accuracy is selected for training on the training data set to obtain the decision function. The prediction results in the decision function are compared with the true labels of the test data set, and the second SVM classifier model that has completed the training is output.
[0071] The loss function of the second SVM classifier model adopts RMSE, and its calculation formula is:
[0072]
[0073] Among them, y′ i is the actual power value, is the predicted power value.
[0074] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A wind power ultra-short-term power prediction method based on machine learning, characterized in that: The steps of the method include: Obtain all-sky image data and power data of wind farms; Perform image processing on all-sky image data of wind farms; Extracting features from the wind farm full sky image data after image processing, and normalizing the features of the wind farm full sky image data; Constructing a first SVM classifier model, taking the features of the wind farm full sky image data as input of the first SVM classifier model, outputting the classification results of the wind farm full sky image data, and screening and outputting the cloud mask image data; The power characteristics of the wind farm are calculated based on the cloud mask image data, a second SVM classifier model is constructed, the power characteristics of the wind farm are used as input of the second SVM classifier model, and the predicted power value of the wind farm is output.
2. The method for wind power ultra-short-term power prediction based on machine learning according to claim 1 is characterized in that: The image processing of the full sky image data of the wind farm is specifically as follows: the full sky image data of the wind farm is synchronously acquired based on multiple adjacent devices, the Euclidean distance of the RGB channel histogram between the devices is calculated, the device whose Euclidean distance exceeds the set value is identified as an abnormal device, and the histogram equalization processing is performed on the detected abnormal device image, and the image processing of the full sky image data of the wind farm is completed through multiple iterations of optimization and correction.
3. The method for wind power ultra-short-term power prediction based on machine learning according to claim 1 is characterized in that: The characteristics of the wind farm full sky image data include: RGB three-channel value characteristics, red-to-blue ratio characteristics, Gaussian spatial filtering characteristics based on a set size window, and brightness channel standard deviation characteristics within a set size window.
4. The method for wind power ultra-short-term power prediction based on machine learning according to claim 3 is characterized in that: The training process of the first SVM classifier model is: Obtain historical all-sky image data of the wind farm and the corresponding manually labeled cloud / sky pixel labels, and perform image processing and feature extraction; The features and corresponding labels of the historical all-sky image data of the wind farm are divided into a training set and a test set; Construct the first SVM classifier model, define the kernel function, optimization objectives and constraints of the first SVM classifier model, set the parameter search range, divide the training set into 5 parts, and cycle each training subset 5 times, with 1 part for validation and the other 4 parts for training; Calculate the verification accuracy and take the average accuracy of 5 times as the performance indicator of this group of parameters. Select the parameter combination with the highest average accuracy to train on the training set to obtain the decision function. Compare the prediction results in the decision function with the true labels of the test set and output the first SVM classifier model that has completed the training.
5. The method for wind power ultra-short-term power prediction based on machine learning according to claim 4 is characterized in that: The kernel function of the first SVM classifier model is calculated as follows: K(x i ,x j )=exp(-γ‖x i -x j ‖ 2 ) The optimization objective and constraint conditions of the first SVM classifier model are calculated as follows: 0≤α i ≤C,i=1,...,n Among them, K(x,x') is the RBF kernel function, α i , α j are the Lagrange multipliers of the i-th and j-th sample points respectively, α is the Lagrange multiplier set, x i is the feature vector of the i-th training sample, γ is the kernel function parameter, y i is the label of the i-th sample, n is the total number of samples, and C is the penalty parameter.
6. The method for wind power ultra-short-term power prediction based on machine learning according to claim 5 is characterized in that: The decision function is calculated as follows: Where b is the bias.
7. The method for wind power ultra-short-term power prediction based on machine learning according to claim 6 is characterized in that: The training process of the second SVM classifier model is: The historical full-sky image data of the wind farm is used for cloud detection through the first SVM classifier model to obtain historical cloud mask image data. The power characteristics of the wind farm are calculated based on the historical cloud mask image data: cloud coverage, cloud motion characteristics, cloud type distribution, and the historical power values at the corresponding time are collected. After preprocessing, the power characteristics and power values of the wind farm are divided into a training data set and a test data set in chronological order; Set the parameter search range, divide the training data set into 5 parts evenly, and cycle through each training data subset 5 times, with 1 part for validation and the other 4 for training; The verification accuracy is calculated, and the average accuracy of 5 times is taken as the performance indicator of this group of parameters. The parameter combination with the highest average accuracy is selected for training on the training data set to obtain the decision function. The prediction results in the decision function are compared with the true labels of the test data set, and the second SVM classifier model that has completed the training is output.
8. The method for wind power ultra-short-term power prediction based on machine learning according to claim 7 is characterized in that: The loss function of the second SVM classifier model adopts RMSE, and its calculation formula is: Among them, y′ i is the actual power value, is the predicted power value.