Method, device and equipment for training of wind power prediction model and power prediction
By combining singular spectrum analysis and a distributed computing framework with an SVM model, the problem of a single deep learning model being unable to analyze complex multimodal wind power sequences is solved, thereby improving the accuracy and computational efficiency of wind power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA TELECOM CLOUD TECH CO LTD
- Filing Date
- 2022-08-19
- Publication Date
- 2026-05-22
AI Technical Summary
Existing single deep learning models struggle to analyze complex, multimodal wind power sequences, resulting in low wind power prediction accuracy and the inability of single-machine prediction systems to handle the massive computational load.
Singular spectral analysis is introduced to decompose wind power sequences. A distributed computing framework is used to accelerate the calculation of sub-models. The support vector machine (SVM) model is combined to optimize the modal prediction results. A temporal convolutional network (TCN) model is used to extract single-modal feature vectors. The system is then used for training and prediction.
It improves the accuracy of wind power prediction, reduces the consumption of computing resources, adapts to the dynamic changes of complex wind environments, and achieves more efficient prediction results.
Smart Images
Figure CN115358314B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the application of machine learning in the field of wind power, and more particularly to methods, apparatus and equipment for training wind power prediction models and predicting power. Background Technology
[0002] With the increasing depletion of fossil fuels, humanity urgently needs an inexhaustible alternative energy source, and wind power is one such option. Predicting wind power output in a region can reveal its wind energy reserves, serving as a crucial reference for wind farm investment. In the field of power dispatching, wind power generation exhibits severe short-term oscillations. Therefore, to mitigate the risks of grid instability caused by wind power generation, a system capable of accurately predicting wind power output is needed, enabling relevant departments to effectively manage reserve capacity and generator operation plans.
[0003] In recent years, deep learning technology has made rapid progress in the field of wind power forecasting. Deep learning uses a large amount of data to train models to extract the characteristic modes of wind power sequences, and makes accurate predictions on each mode. Then, the prediction results are weighted and summed to form the final prediction value.
[0004] However, existing single deep learning models struggle to analyze complex, multimodal wind power sequences. Therefore, a hybrid model has emerged that pre-decomposes the wind power sequence into sub-sequences and then uses a deep learning model to perform short-term predictions on sub-sequences with similar modes, such as the SSA-TCN model. While this hybrid model can set a prediction model for each mode of wind power, greatly improving prediction accuracy, wind power sequences typically decompose into dozens of modes, and a single-machine prediction system cannot handle such a massive computational load. Summary of the Invention
[0005] This invention provides a wind power prediction model method, which introduces singular spectrum analysis to thoroughly analyze wind power sequences, then uses a distributed computing framework to accelerate the calculation of sub-models, and finally uses a support vector machine (SVM) model to optimize the modal prediction results, thus solving the technical problem of low prediction accuracy in existing wind power prediction methods when dealing with complex wind conditions.
[0006] Firstly, this application provides a method for training a wind power prediction model, comprising:
[0007] Obtain the original wind power sequence corresponding to different historical sampling points;
[0008] The original wind power sequence is subjected to singular value decomposition, and multiple wind power subsequences corresponding to different modes are decomposed from the original wind power sequence according to the energy contribution value corresponding to different singular values.
[0009] A sliding window is used to extract data from the wind power subsequences of each mode to obtain window data for each mode, and a training set is divided from the window data of each mode.
[0010] A distributed computing approach is adopted, in which window data from each modality training set is input into the Temporal Convolutional Network (TCN) model for single-modality feature vector extraction;
[0011] The SVM prediction model is used to predict power in a single mode based on the feature vectors of each single mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function in each mode.
[0012] In one or more embodiments, based on the energy contribution values corresponding to different singular values, multiple wind power sub-sequences corresponding to different modes are decomposed from the original wind power sequence, including:
[0013] The singular values are sorted in descending order of their energy contribution to obtain the singular spectrum;
[0014] The singular values corresponding to the noise component are determined from the singular spectrum, and the power corresponding to the singular values of the noise component is removed from the original wind power sequence.
[0015] For the singular values in the non-noise part of the singular spectrum, the singular values whose energy contribution values are within a set floating range are taken as the power values in the original wind power sequence and used as a wind power subsequence corresponding to a mode.
[0016] In one or more embodiments, for singular values in the non-noise portion of the singular spectrum, the singular values whose energy contribution values are within a set floating range are taken as the power values in the original wind power sequence, and treated as a wind power sub-sequence corresponding to a mode, including:
[0017] The triples obtained from singular values are arranged according to The component matrix is restored to the trajectory matrix;
[0018] The resulting component matrix is averaged along the secondary diagonal using the following formula to obtain a subsequence of the wind power sequence:
[0019]
[0020] Where N is the length of the original wind power sequence, and b s,t It is matrix X′ i The element in the s-th row and t-th column, X′ i (j) is the component matrix X′ i The value of the j-th sample point after serialization.
[0021] In one or more embodiments, when this method is applied to a main control computer, a distributed computing approach is used to input window data from each modality training set into a Temporal Convolutional Network (TCN) model for single-modality feature vector extraction, including:
[0022] The main control unit will divide the training set from the window data of each modality and distribute it to each sub-machine. The sub-machine will input the window data in the corresponding modality training set into the TCN model to extract the single modality feature vector and report the extracted single modality feature vector to the main control unit.
[0023] The SVM prediction model is used to predict power in a single mode based on the feature vectors of each mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function for each mode, including:
[0024] The main control unit uses the SVM prediction model to predict the power of each mode based on the feature vector of each mode, determines the loss function of each mode, generates the transfer gradient based on the loss function of each mode, and sends it to the corresponding sub-machine. Each machine updates the TCN model parameters based on the transfer gradient.
[0025] In one or more embodiments, a sliding window is used to retrieve data for each mode of wind power subsequence, including:
[0026] Determine the length and step size of the sliding window;
[0027] The sliding step size is used to slide on the wind power subsequence of each mode. Each time the sliding step is used, the data of the sliding length is taken as a sample.
[0028] The obtained samples are divided into training set and test set.
[0029] In one or more embodiments, the method further includes:
[0030] The window data from each modal test set is input into the corresponding modal TCN to extract modal feature vectors. The SVM prediction model is used to perform single-step prediction based on the extracted feature vectors. The accuracy of the single-step prediction is determined. The power of the single-step prediction is combined with the power of the previous sampling points to form new data of a sliding window length. The new data is input into the TCN model, and the SVM model is used to perform single-step prediction again. The above process is repeated until the prediction of the set multi-step length is obtained, and the prediction accuracy of the set multi-step length is determined.
[0031] Once the prediction accuracy for both single-step and multi-step predictions meets the accuracy requirements, training ends, and the optimal prediction step size for the wind power prediction model is determined based on the prediction accuracy for both single-step and multi-step predictions.
[0032] Secondly, this application provides a method for power prediction using a wind power prediction model, comprising:
[0033] Obtain the original wind power sequence corresponding to different historical sampling points;
[0034] The original wind power sequence is subjected to singular value decomposition, and multiple wind power subsequences corresponding to different modes are decomposed from the original wind power sequence according to the energy contribution value corresponding to different singular values.
[0035] For each mode of wind power subsequence, a sliding window is used to take the window data of the most recent sampling time;
[0036] The window data of each mode is input into the TCN model to extract the single-mode feature vector. The SVM prediction model is then used to predict the power of future sampling points under each single mode based on the single-mode feature vector.
[0037] By using the predicted power of future sampling points under different modes, the wind power sequence is reconstructed according to the inverse process of singular value decomposition.
[0038] Thirdly, this application provides an apparatus for training a wind power prediction model, comprising:
[0039] The data processing module is used to obtain the original wind power sequence corresponding to different historical sampling points;
[0040] The subsequence determination module is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values.
[0041] The window partitioning module is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode.
[0042] The feature vector extraction module is used to extract single-modal feature vectors by inputting window data from each modality training set into the Temporal Convolutional Network (TCN) model using a distributed computing approach.
[0043] The parameter adjustment module is used to perform power prediction in a single mode using the SVM prediction model based on the feature vectors of each single mode, and to adjust the parameters of the SVM prediction model and the corresponding TCN model based on the loss function in each mode.
[0044] In one or more embodiments, the subsequence determination module decomposes multiple wind power subsequences corresponding to different modes from the original wind power sequence based on the energy contribution values corresponding to different singular values, including:
[0045] The singular values are sorted in descending order of their energy contribution to obtain the singular spectrum;
[0046] The singular values corresponding to the noise component are determined from the singular spectrum, and the power values corresponding to the singular values of the noise component are removed from the original wind power power sequence.
[0047] For the singular values in the non-noise part of the singular spectrum, the power values in the original wind power sequence corresponding to the singular values whose energy contribution values are within a set floating range are taken as a wind power subsequence corresponding to a mode.
[0048] In one or more embodiments, the subsequence determination module is further configured to, for singular values in the non-noise portion of the singular spectrum, take the singular values whose energy contribution values are within a set floating range as the power values in the original wind power sequence, as a wind power subsequence corresponding to a mode, including:
[0049] The triples obtained from the singular values are restored to the component matrix of the trajectory matrix;
[0050] The resulting component matrix is averaged along the secondary diagonal using the following formula to obtain a subsequence of the wind power sequence:
[0051]
[0052] Where N is the length of the original wind power sequence, is the element in the s-th row and t-th column of the matrix, and is the value of the j-th sampling point after the component matrix is serialized.
[0053] In one or more embodiments, when the method is applied to the main control unit, the feature vector extraction module inputs window data from each modality training set into the Temporal Convolutional Network (TCN) model to extract single-modality feature vectors, including:
[0054] The main control unit will divide the training set from the window data of each modality and distribute it to each sub-machine. The sub-machine will input the window data in the corresponding modality training set into the TCN model to extract the single modality feature vector and report the extracted single modality feature vector to the main control unit.
[0055] The SVM prediction model is used to predict power in a single mode based on the feature vectors of each mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function for each mode, including:
[0056] The main control unit uses the SVM prediction model to predict the power of each mode based on the feature vector of each mode, determines the loss function of each mode, generates the transfer gradient based on the loss function of each mode, and sends it to the corresponding sub-machine. Each machine updates the TCN model parameters based on the transfer gradient.
[0057] In one or more embodiments, the window partitioning module uses a sliding window to retrieve data for each mode of wind power sub-sequence, including:
[0058] Determine the length and step size of the sliding window;
[0059] The sliding step size is used to slide on the wind power subsequence of each mode. Each time the sliding step is used, the data of the sliding length is taken as a sample.
[0060] The obtained samples are divided into training set and test set.
[0061] In one or more embodiments, it further includes:
[0062] The optimal step size determination module is used to extract modal feature vectors from the window data input of each modality test set into the corresponding modality TCN model, use the SVM prediction model to perform single-step prediction based on the extracted feature vectors, determine the accuracy of single-step prediction, recombine the power of single-step prediction with the power of the previous sampling points into new data of a sliding window length, input the new data into the TCN model, and use the SVM model to perform single-step prediction again. The above process is repeated until the prediction of the set multi-step length is obtained, and the prediction accuracy of the set multi-step length is determined.
[0063] Once the prediction accuracy for both single-step and multi-step predictions meets the accuracy requirements, training ends, and the optimal prediction step size for the wind power prediction model is determined based on the prediction accuracy for both single-step and multi-step predictions.
[0064] Fourthly, this application provides an apparatus for power prediction of a wind power prediction model, comprising:
[0065] The data processing module is used to obtain the original wind power sequence corresponding to different historical sampling points;
[0066] The subsequence determination module is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values.
[0067] The window partitioning module is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode.
[0068] The prediction module is used to input window data of each modality into the TCN model for single-modal feature vector extraction using a distributed computing approach, and then use the SVM prediction model to predict the power of future sampling points under each single-modal feature vector.
[0069] The reconstruction module reconstructs the wind power sequence by using the predicted power of future sampling points under different modes and following the inverse process of singular value decomposition.
[0070] Fifthly, this application provides an apparatus for training and predicting wind power forecasting models, the apparatus comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method for training the wind power forecasting model, or to perform the method for executing the wind power forecasting model.
[0071] The wind power prediction model method provided in this application has the following beneficial effects:
[0072] This paper introduces singular spectrum analysis to thoroughly analyze wind power sequences, utilizes a distributed computing framework to accelerate sub-model calculations, and finally uses a support vector machine (SVM) model to optimize the modal prediction results. This solves the technical problem of low prediction accuracy in existing wind power prediction methods when dealing with complex wind conditions. Attached Figure Description
[0073] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0074] Figure 1 A flowchart illustrating the training method for the wind power prediction model provided in this application embodiment;
[0075] Figure 2 Singular spectra provided for embodiments of this application;
[0076] Figure 3 A diagram of a prediction model based on the distributed computing framework SSA-TCN-SVM provided in this application embodiment;
[0077] Figure 4 A structural diagram of the Temporal Convolutional Network (TCN) model provided in the embodiments of this application;
[0078] Figure 5 A flowchart illustrating the overall training method for the wind power prediction model provided in this application embodiment;
[0079] Figure 6 A comparison chart of performance indicators with existing technologies provided for embodiments of this application;
[0080] Figure 7 A flowchart of the wind power prediction model method provided in the embodiments of this application;
[0081] Figure 8This is a schematic diagram of the overall wind power prediction model and method provided in the embodiments of this application;
[0082] Figure 9 A schematic diagram of a device for training a wind power prediction model provided in an embodiment of this application;
[0083] Figure 10 A schematic diagram of a power prediction device for a wind power prediction model provided in an embodiment of this application;
[0084] Figure 11 A schematic diagram of the equipment for training and power prediction of the wind power prediction model provided in the embodiments of this application. Detailed Implementation
[0085] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0086] Existing single deep learning models struggle to analyze complex, multimodal wind power sequences, leading to the development of hybrid models. Hybrid models involve pre-decomposing the wind power sequence into sub-sequences, and then using a deep learning model to perform short-term predictions on sub-sequences with similar modes, such as the SSA-TCN model. This approach significantly improves the algorithm's prediction accuracy. However, wind power sequences typically decompose into dozens of modes. Theoretically, setting up a model for each mode could perfectly analyze all wind power sequences. However, full-modal prediction consumes enormous computational resources, making it unsuitable for single-machine prediction systems.
[0087] In view of the above problems, this application provides a training method for a wind power prediction model, such as... Figure 1 As shown, the method includes:
[0088] Step S101: Obtain the original wind power sequence corresponding to different historical sampling points;
[0089] To obtain historical operating data from a wind farm, since there are multiple wind turbines, it is necessary to merge the historical power data of each turbine. Specifically, this includes adding and merging data from the same sampling point, determining the data corresponding to different sampling points according to the sampling order in the time dimension, filling in missing values for sampling points with missing values (such as zero, mean, or median), filtering out sampling points with abnormal data, and setting the power value of sampling points with abnormal data to 0. The processed results are then stored on the distributed file system HDFS.
[0090] Step S102: Perform singular value decomposition on the original wind power sequence, and decompose multiple wind power subsequences corresponding to different modes from the original wind power sequence according to the energy contribution value corresponding to different singular values.
[0091] In one or more possible embodiments, the singular values decomposed from the original wind power sequence are sorted in descending order of energy contribution value to obtain a singular spectrum as follows: Figure 2 The horizontal axis represents 74 singular values, and the vertical axis represents the energy contribution value. The tail of the energy contribution value that is lower than a set threshold is the noise part. The power value corresponding to the singular value corresponding to the noise part is removed from the original wind power power sequence.
[0092] For the singular values corresponding to the non-noise portion of the singular spectrum, the power values in the original wind power sequence corresponding to the singular values whose energy contribution values are within a set floating range are taken as a wind power subsequence corresponding to a mode.
[0093] The specific method for obtaining the power subsequence is given below:
[0094] (1) Embedding: Extract a sliding window vector of length L from the wind power sequence and embed it into the trajectory matrix as a column vector according to formula ①.
[0095]
[0096] Among them, a ij Let x represent the data in the i-th row and j-th column of the matrix. i This represents the power at the i-th sampling point in the wind power power sequence.
[0097] (2) Use the Schmidt orthogonal method to solve for the eigenvectors of all input matrices step by step, and then solve for the left and right eigenvectors of the singular value decomposition according to certain rules. Perform singular value decomposition according to formula ② to obtain triples.
[0098]
[0099] The above method for solving the left and right eigenvector matrices U and V using singular value decomposition is the existing approach and will not be detailed here. M is a diagonal matrix.
[0100] (3) Restore the triples obtained from the singular values to the component matrix of the trajectory matrix according to formula ③; then take the average of the component matrix along the secondary diagonal according to formula ④ to obtain the subsequence of the wind power sequence:
[0101]
[0102]
[0103] Where N is the length of the original wind power sequence, and b s,t It is matrix X′ i The element in the s-th row and t-th column, X′ i (j) is the component matrix X′ i After serialization, the value of the j-th sampling point is the same as the value of u mentioned above. i It is an eigenvector in the left eigenvector matrix, v i These are the eigenvectors in the right eigenvector matrix. Since the subsequences of the wind power sequence are decomposed from the original power sequence, the corresponding wind power sequence can be reconstructed using the subsequences, specifically using the following formula:
[0104] X(t)=X′1(t)+X′2(t)+…+X′ r (t) ⑤
[0105] During the model training phase, it is not necessary to reconstruct the wind power sequence; instead, sub-model training is performed based on the decomposed wind power sub-sequences.
[0106] Step S103: Take data from the wind power subsequence of each mode using a sliding window to obtain the window data of each mode, and divide the training set from the window data of each mode;
[0107] Based on the historical data sampling points used, the window step size for data collection for each mode's corresponding subsequence is determined. Each wind power subsequence is windowed, that is, the elements in the sequence are replaced with a window of length L, and the next window of data is collected by sliding according to the window step size, resulting in multiple window data. These multiple window data are used as a sample set to divide the test set and the training set. For example, the training set and the test set can be divided into a 70% ratio.
[0108] Taking a sliding step of 1 sampling point as an example, the power of the first to L sampling points of the wind power subsequence is taken as a window of data, i.e., as a sample. The power of the second to L+1 sampling points is taken as a sample, and so on to obtain other samples.
[0109] In one or more possible embodiments, a sliding window is used to retrieve data for each mode of wind power subsequence, including:
[0110] Determine the length and step size of the sliding window;
[0111] The sliding step size is used to slide on the wind power subsequence of each mode. Each time the sliding step is used, the data of the sliding length is taken as a sample.
[0112] The obtained samples are divided into training set and test set.
[0113] Step S104: Using a distributed computing approach, the window data from each modality training set is input into the Temporal Convolutional Network (TCN) model for single-modality feature vector extraction.
[0114] The TCN model consists of an input layer, a feature vector extraction layer, and an output layer. The input layer takes in window data, the feature extraction layer extracts short-term features from the window data of a single modality, and the output layer outputs the corresponding single-modality feature vector.
[0115] Step S105: Use the SVM prediction model to predict the power of each single mode based on the feature vector of each single mode, and adjust the parameters of the SVM prediction model and the corresponding TCN model based on the loss function of each mode.
[0116] This application utilizes an SVM prediction model to predict the power of future sampling points based on single-modal feature vectors. For each TCN model, a loss function is determined based on the predicted power and the actual power. Based on the loss function, the adjustment parameters of the corresponding TCN model and the SVM prediction model are determined.
[0117] This application's embodiments are based on a distributed system design algorithm. Wind power sequences are decomposed using a time series decomposition algorithm, followed by feature extraction via deep learning. Finally, the prediction results of all sub-models are fitted using a regression algorithm to obtain a highly accurate final result. This application performs singular spectrum decomposition on complex, multimodal, highly random, and highly uncertain wind power sequences, leveraging the high computing power of distributed computing clusters to thoroughly decompose wind power. Simultaneously, this application introduces a time convolutional network based on SSA-TCN-SVM, which can fully extract all short-term features of the wind power sequence, allowing the predicted curve to express more details. Furthermore, by introducing the convolutional network, each element of the input sequence undergoes the same convolution operation to obtain the final result, significantly improving the prediction lag. This application uses a distributed training method for SSA-TCN-SVN, simultaneously performing parallel computation to accelerate matrix decomposition, subsequence generation, and training speeds. Utilizing a distributed scenario expands the number of models, resulting in more accurate prediction results. This method also employs SVM to optimize the output sequences of each sub-model, expands the weighting factor of the main power factor sequence, and suppresses the minor components that are not major, thereby reducing the randomness and uncertainty of the wind power sequence and finding the true trend term and the main detailed components.
[0118] In one or more possible embodiments, when the method is applied to the main control machine, a distributed computing approach is used to input window data from each modality training set into the Temporal Convolutional Network (TCN) model for single-modality feature vector extraction, including:
[0119] The main control unit will divide the training set from the window data of each modality and distribute it to each sub-machine. The sub-machine will input the window data in the corresponding modality training set into the TCN model to extract the single modality feature vector and report the extracted single modality feature vector to the main control unit.
[0120] The SVM prediction model is used to predict power in a single mode based on the feature vectors of each mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function for each mode, including:
[0121] The main control unit uses the SVM prediction model to predict the power of each mode based on the feature vector of each mode, determines the loss function of each mode, generates the transfer gradient based on the loss function of each mode, and sends it to the corresponding sub-machine. Each machine updates the TCN model parameters based on the transfer gradient.
[0122] The prediction model based on the distributed computing framework SSA-TCN-SVM provided in this embodiment Figure 3 As shown, each machine contains a TCN sub-model, where TCN stands for Temporal Convolutional Network.
[0123] TCN can extract short-term features from wind power sequences and output the extracted feature vectors to the main controller.
[0124] Each slave machine has two functions: forward and backward. Forward is used to generate the final output based on the mode, and backward is used to obtain the transmitted gradient from the master machine and update the parameters of the sub-model based on it.
[0125] The master control machine is used to generate the final output and loss function. Based on the loss function result, it generates and broadcasts the gradient to each machine. The loss function value is compared with a threshold, and an early stop strategy is set to terminate training. Because wind conditions are constantly changing—for example, the model will differ between winter and summer—to enable the distributed computing framework to dynamically update the model, the master control machine is used for long-cycle model training. Each long cycle involves phased model training and the release of official model predictions. However, within the smaller cycles of a long cycle, the master control machine maintains its operating parameters in normal operating mode, while the pre-training master control machine is in training mode, continuously backing up historical data and controlling the slave machines to train during idle times. Once the prediction accuracy drops below a threshold, the parameters are updated. This two-cycle training fully utilizes the machines and is adaptable to ever-changing wind conditions, dynamically adjusting the model.
[0126] In this application's embodiments, single-step prediction is performed during model training, such as... Figure 4The diagram shows the structure of the Temporal Convolutional Network (TCN) model. This TCN model consists of one-dimensional causal convolutions, thus exhibiting temporal characteristics. The TCN model combines residual networks and dilated convolutions to significantly reduce the network depth required for the same input sequence, thereby mitigating the impact of gradient vanishing.
[0127] During the training process of this application, the TCN model is combined with the SVM model for single-step prediction, predicting the data at the next future sampling point, such as... Figure 4 As shown, input the data at the 1st to nth sampling points, and output the data at the 2nd to (n+1)th sampling points.
[0128] In one or more possible embodiments, the method further includes:
[0129] The window data from each modal test set is input into the corresponding modal TCN model to extract single-modal feature vectors. The SVM prediction model is used to perform single-step prediction based on the extracted single-modal feature vectors. The accuracy of the single-step prediction is determined. The power of the single-step prediction is combined with the power of the previous sampling points to form new data of a sliding window length. The new data is input into the TCN model, and the SVM model is used to perform single-step prediction again. The above process is repeated until the prediction of the set multi-step length is obtained, and the prediction accuracy of the set multi-step length is determined.
[0130] Once the prediction accuracy for both single-step and multi-step predictions meets the accuracy requirements, training ends, and the optimal prediction step size for the wind power prediction model is determined based on the prediction accuracy for both single-step and multi-step predictions.
[0131] In step S105 above, this embodiment of the application replaces the DenseLayer layer used in the traditional TCN model with the SVM model optimized by the support vector machine. By using the SVM model vector optimization theory to predict the power of future sampling points, better prediction results can be obtained.
[0132] The optimization theory of SVM is explained below. The symbols involved in the formulas can be determined using relevant techniques and will not be elaborated upon here. The optimization problem of Support Vector Machine can be expressed by the following formula:
[0133]
[0134] The dual problem of this optimization problem:
[0135]
[0136] Output layer parameter update:
[0137]
[0138]
[0139] This paper employs SVM (Single Vector Machine) model vector optimization theory to predict the power of future sampling points. By transforming the linear regression problem into a nonlinear problem based on the kernel function space, it better separates the boundaries of each sample point. Using the SVM model yields better generalization, enabling our trained wind power prediction model to withstand more complex wind power scenarios and obtain more accurate prediction results.
[0140] like Figure 5 This is an overall flowchart of the training method for the wind power prediction model in the embodiments of this application, which specifically includes:
[0141] Step S501: Obtain the original wind power sequence corresponding to different historical sampling points;
[0142] Step S502: Preprocess the historical values of the wind power series, fill missing values with zero, mean or median, and set outliers to 0;
[0143] Step S503: Perform singular spectrum decomposition on the preprocessed wind power sequence, remove noise based on the singular spectrum, and decompose the non-noise part into multiple wind power subsequences corresponding to the modes.
[0144] The singular value decomposition of the preprocessed wind power sequence is used to obtain the singular spectrum. Based on the singular spectrum, the noise part of the wind power sequence with an energy contribution value lower than a set threshold is removed, and the non-noise part is decomposed into multiple wind power subsequences corresponding to modes.
[0145] Step S504: Take data from the wind power subsequence of each mode using a sliding window, obtain the window data of each mode, and divide it into training set and test set;
[0146] Step S505: Input the window data from each modality training set into the Temporal Convolutional Network (TCN) model to extract single-modality feature vectors and send them to the main control computer.
[0147] Step S506: The main control machine uses the SVM prediction model to predict the power of each single mode based on the feature vector of each single mode, calculates the loss function and generates the gradient to be transmitted to the corresponding sub-machine, and adjusts the parameters of the SVM prediction model and the corresponding TCN model.
[0148] Step S507: Each machine updates the TCN model parameters according to the transmitted gradient. Step S508 is executed when the main control machine confirms that the TCN model has converged.
[0149] Step S508: Input the window data from each modal test set into the corresponding modal TCN model to extract modal feature vectors and send them to the main control computer;
[0150] Step S509: The main control unit uses the SVM prediction model to perform single-step prediction based on the extracted feature vectors and determines the accuracy of the single-step prediction.
[0151] Step S510: Combine the power predicted in a single step with the power of the previous sampling points into new data of a sliding window length. Input the new data into the TCN model and use the SVM prediction model to perform single-step prediction again. Repeat the above process until the prediction with the set multi-step length is obtained, and determine the prediction accuracy of the set multi-step length.
[0152] Step S511: When the prediction accuracy of single step and set multi-step is determined to meet the accuracy requirements, the training ends, and the optimal prediction step size of the wind power prediction model is determined based on the prediction accuracy of single step and set multi-step.
[0153] like Figure 6 As shown, the method proposed in this invention has a significant performance improvement compared to the comparative method.
[0154] It can be seen that SSA-TCN-SVM has the lowest RMSE and the highest R², indicating that SSA-TCN-SVM fits the true curve well, has the smallest prediction error oscillation, and the highest prediction accuracy. Compared with SSA-TCN-SVM, the TCN model has the fastest exponential decrease in R², while SSA-TCN-SVM has the slowest exponential decrease in R².
[0155] This application also provides a method for power prediction using a wind power prediction model, such as... Figure 7 As shown, it includes:
[0156] Step 701: Obtain the original wind power sequence corresponding to different historical sampling points;
[0157] Step 702: Perform singular value decomposition on the original wind power sequence, and decompose multiple wind power subsequences corresponding to different modes from the original wind power sequence according to the energy contribution value corresponding to different singular values.
[0158] Step 703: For each mode of wind power subsequence, a sliding window is used to take the window data of the most recent sampling time;
[0159] Step 704: Input the window data of each modality into the TCN model to extract the single-modal feature vector, and use the SVM prediction model to predict the power of future sampling points under each single-modal feature vector.
[0160] Step 705: Using the predicted power of future sampling points under different modes, reconstruct the wind power sequence according to the inverse process of singular value decomposition.
[0161] When using wind power prediction models to predict the power of future sampling points, the power of a single sampling point can be predicted, or the power of sampling points with the optimal prediction step size can be predicted, or the power of sampling points over a period of time can be predicted according to requirements. This is achieved by combining the newly predicted data with the previous data and inputting it into the model. For details, please refer to the above description, which will not be repeated here.
[0162] like Figure 8 The diagram shows the overall wind power prediction model method. The method of performing singular value decomposition on the original wind power sequence is described in the above embodiment. The obtained wind power subsequences are assigned to corresponding submachines. Each submachine uses the TCN model to extract single-mode feature vectors. The SVM prediction model is used to predict the power of future sampling points under each single-mode feature vector. New inputs are continuously generated to predict the data of subsequent sampling points. The prediction results are sent to the main control unit. The SVM regression prediction model of the main control unit determines the wind power subsequences of sampling points in different modes within a preset time period and reconstructs them into the predicted wind power sequence according to the inverse process of singular value decomposition.
[0163] Based on the same inventive concept, this application also provides a device for training a wind power prediction model, such as... Figure 9 As shown, it includes:
[0164] Data processing module 901 is used to obtain the original wind power sequence corresponding to different historical sampling points;
[0165] The subsequence determination module 902 is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values.
[0166] The window partitioning module 903 is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode.
[0167] The feature vector extraction module 904 is used to extract single-modal feature vectors by inputting window data from each modality training set into the temporal convolutional network (TCN) model using a distributed computing approach.
[0168] The parameter adjustment module 905 is used to perform power prediction in a single mode using the SVM prediction model based on the feature vectors of each single mode, and to adjust the parameters of the SVM prediction model and the corresponding TCN model based on the loss function in each mode.
[0169] Based on the same inventive concept, this application also provides a device for power prediction of a wind power prediction model, such as... Figure 10 As shown, it includes:
[0170] Data processing module 1001 is used to obtain the original wind power sequence corresponding to different historical sampling points;
[0171] The subsequence determination module 1002 is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values.
[0172] The window partitioning module 1003 is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode.
[0173] The prediction module 1004 is used to input window data of each mode into the TCN model for single-mode feature vector extraction using a distributed computing method, and use the SVM prediction model to predict the power of future sampling points under each single mode based on the single-mode feature vector.
[0174] The reconstruction module 1005 reconstructs the wind power sequence by using the predicted power of future sampling points under different modes and following the inverse process of singular value decomposition.
[0175] Based on the same inventive concept, this application also provides a device 1100 for training a wind power prediction model and predicting power, such as... Figure 11 As shown, it includes at least one processor 1102; and a memory 1001 communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the training method of the wind power prediction model described above, or to execute the method of the wind power prediction model.
[0176] Memory 1101 is used to store programs. Specifically, the program may include program code, which includes computer operation instructions. Memory 1101 may be volatile memory, such as random-access memory (RAM); it may also be non-volatile memory, such as flash memory, hard disk drive (HDD), or solid-state drive (SSD); or it may be any one or a combination of the above-mentioned volatile and non-volatile memory types.
[0177] Processor 1102 can be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. It can also be a hardware chip. The aforementioned hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The aforementioned PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.
[0178] This disclosure also provides a computer program product that, when run on an electronic device, causes the electronic device to execute any of the wind power prediction model methods described above or any possible method involved in any of the wind power prediction model methods.
[0179] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, apparatuses, or modules, and may be electrical, mechanical, or other forms.
[0180] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0181] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.
[0182] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.
[0183] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from a website site, computer, server, or data center to another website site, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).
[0184] The technical solutions provided in this application have been described in detail above. Specific examples have been used in this application to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
[0185] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0186] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood by those skilled in the art from the computer program instructions that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of one or more computer-usable storage media containing computer-usable program code (implementing each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams). These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0187] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0188] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0189] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A training method for a wind power prediction model, characterized in that, include: Obtain the original wind power sequence corresponding to different historical sampling points; The original wind power sequence is subjected to singular value decomposition, and multiple wind power subsequences corresponding to different modes are decomposed from the original wind power sequence according to the energy contribution value corresponding to different singular values. A sliding window is used to extract data from the wind power subsequences of each mode to obtain window data for each mode, and a training set is divided from the window data of each mode. A distributed computing approach is adopted, in which window data from each modality training set is input into the Temporal Convolutional Network (TCN) model for single-modality feature vector extraction; The SVM prediction model is used to predict power in a single mode based on the feature vectors of each single mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function in each mode.
2. The method as described in claim 1, characterized in that, Based on the energy contribution values corresponding to different singular values, multiple wind power sub-sequences corresponding to different modes are decomposed from the original wind power sequence, including: The singular values are sorted in descending order of their energy contribution to obtain the singular spectrum; The singular values corresponding to the noise component are determined from the singular spectrum, and the power values corresponding to the singular values of the noise component are removed from the original wind power power sequence. For the singular values in the non-noise part of the singular spectrum, the power values in the original wind power sequence corresponding to the singular values whose energy contribution values are within a set floating range are taken as a wind power subsequence corresponding to a mode.
3. The method as described in claim 2, characterized in that, For singular values in the non-noise portion of the singular spectrum, the singular values whose energy contribution values are within a set fluctuation range are taken as the power values in the original wind power sequence, and treated as a wind power subsequence corresponding to a mode, including: The triples obtained from singular values are arranged according to The component matrix restored to the trajectory matrix, where u i It is an eigenvector in the left eigenvector matrix, v i These are the eigenvectors in the right eigenvector matrix. It is the i-th singular value; The resulting component matrix is averaged along the secondary diagonal using the following formula to obtain a subsequence of the wind power sequence: Where N is the length of the original wind power sequence, where, It is a matrix The element in the s-th row and t-th column, It is a component matrix After serialization, the value of the j-th sampling point, where L is the length of the sliding window vector.
4. The method as described in claim 1, characterized in that, When applied to the main control computer, this method employs distributed computing, inputting window data from each modality training set into the Temporal Convolutional Network (TCN) model for single-modality feature vector extraction, including: The main control unit will divide the training set from the window data of each modality and distribute it to each sub-machine. The sub-machine will input the window data in the corresponding modality training set into the TCN model to extract the single modality feature vector and report the extracted single modality feature vector to the main control unit. The SVM prediction model is used to predict power in a single mode based on the feature vectors of each mode. The parameters of the SVM prediction model and the corresponding TCN model are adjusted according to the loss function for each mode, including: The main control unit uses the SVM prediction model to predict the power of each mode based on the feature vector of each mode, determines the loss function of each mode, generates the transfer gradient based on the loss function of each mode, and sends it to the corresponding sub-machine. Each machine updates the TCN model parameters based on the transfer gradient.
5. The method as described in claim 1, characterized in that, A sliding window is used to retrieve data for each mode of wind power subsequence, including: Determine the length and step size of the sliding window; The sliding step size is used to slide on the wind power subsequence of each mode. Each time the sliding step is used, the data of the sliding length is taken as a sample. The obtained samples are divided into training set and test set.
6. The method as described in claim 5, characterized in that, Also includes: The window data from each modal test set is input into the corresponding modal TCN model to extract single-modal feature vectors. The SVM prediction model is used to perform single-step prediction based on the extracted single-modal feature vectors. The accuracy of the single-step prediction is determined. The power of the single-step prediction is combined with the power of the previous sampling points to form new data of a sliding window length. The new data is input into the TCN model, and the SVM model is used to perform single-step prediction again. The above process is repeated until the prediction of the set multi-step length is obtained, and the prediction accuracy of the set multi-step length is determined. Once the prediction accuracy for both single-step and multi-step predictions meets the accuracy requirements, training ends, and the optimal prediction step size for the wind power prediction model is determined based on the prediction accuracy for both single-step and multi-step predictions.
7. A power prediction method for a wind power prediction model, characterized in that, include: Obtain the original wind power sequence corresponding to different historical sampling points; The original wind power sequence is subjected to singular value decomposition, and multiple wind power subsequences corresponding to different modes are decomposed from the original wind power sequence according to the energy contribution value corresponding to different singular values. For each mode of wind power subsequence, a sliding window is used to take the window data of the most recent sampling time; The window data of each mode is input into the TCN model to extract the single-mode feature vector. The SVM model is then used to predict the power of future sampling points under each single mode based on the single-mode feature vector. By using the predicted power of future sampling points under different modes, the wind power sequence is reconstructed according to the inverse process of singular value decomposition.
8. A device for training a wind power prediction model, characterized in that, include: The data processing module is used to obtain the original wind power sequence corresponding to different historical sampling points; The subsequence determination module is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values. The window partitioning module is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode. The feature vector extraction module is used to extract single-modal feature vectors by inputting window data from each modality training set into the Temporal Convolutional Network (TCN) model using a distributed computing approach. The parameter adjustment module is used to perform power prediction in a single mode using the SVM prediction model based on the feature vectors of each single mode, and to adjust the parameters of the SVM prediction model and the corresponding TCN model based on the loss function in each mode.
9. A device for power prediction of a wind power prediction model, characterized in that, include: The data processing module is used to obtain the original wind power sequence corresponding to different historical sampling points; The subsequence determination module is used to perform singular value decomposition on the original wind power sequence and decompose multiple wind power subsequences corresponding to modes from the original wind power sequence according to the energy contribution values corresponding to different singular values. The window partitioning module is used to extract data from the wind power subsequence of each mode using a sliding window, obtain window data for each mode, and partition the training set from the window data of each mode. The prediction module is used to input window data of each modality into the TCN model for single-modal feature vector extraction using a distributed computing approach, and then use the SVM prediction model to predict the power of future sampling points under each single-modal feature vector. The reconstruction module reconstructs the wind power sequence by using the predicted power of future sampling points under different modes and following the inverse process of singular value decomposition.
10. An electronic device for wind power prediction models, characterized in that, The method includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-6, or to perform the method as described in claim 7.