Wind speed data prediction method in wind power generation scene

By converting wind speed data into a three-dimensional matrix and processing with multi-layer perceptron and spatiotemporal encoding modules, the shortcomings of spatial dependence and spatiotemporal data dependence between wind farms are solved, and the accuracy and reliability of wind speed prediction are improved.

CN118966429APending Publication Date: 2024-11-15CHINA UTONE CONSTR CONSULTING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411022552.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-29
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing wind speed prediction methods rarely consider the spatial dependence between wind farms or the limited ability to mine spatial and temporal data dependence, resulting in poor prediction results.

Method used

The wind speed data is converted into a three-dimensional matrix, and the features are extracted through multi-layer perceptron processing and the K-layer spatiotemporal encoding module is processed, including a structured perception attention module and a frequency domain enhanced time delay perception attention module for normalization and math The model is processed to capture periodic information and trend information, and finally generates prediction results through serial operations through the spatiotemporal encoding module.

Benefits of technology

The accuracy and reliability of wind speed prediction are improved, and the understanding and prediction ability of wind speed changes is enhanced by considering multi-dimensional information in time, space and characteristic dimensions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118966429A_ABST
    Figure CN118966429A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of wind power generation and data processing, and discloses a wind speed data prediction method in a wind power generation scene, and the method comprises the steps: reading wind speed data, and converting the wind speed data into a matrix corresponding to three-dimensional wind speed data comprising a time dimension, a space dimension and a feature dimension; processing a three-dimensional matrix corresponding to the obtained wind speed data through a first multi-layer perceptron to obtain first wind speed data; and processing the first wind speed data through a K-layer space-time coding module to obtain to-be-predicted wind speed data in a future time period. According to the invention, through a K-layer space-time coding module, the correlation of wind speed data in time and space is considered, and the wind speed data to be predicted in a future time period is predicted by learning and simulating the relation, so that the trend and the mode of historical wind speed data are considered, and the correlation in space is also integrated; therefore, the prediction result is more accurate and reliable.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation and data processing, and in particular to a method for predicting wind speed data in a wind power generation scenario. Background Art

[0002] In recent years, wind speed prediction in wind power generation scenarios has become one of the key technologies for economic dispatch and safe and stable operation of my country's power system.

[0003] Wind speed is a time series that changes continuously over time. At the same time, the wind speeds between different wind farms have strong spatial dependence. Therefore, wind speed data is a typical spatiotemporal data, and wind speed prediction can be regarded as a typical spatiotemporal prediction problem.

[0004] Spatiotemporal forecasting, represented by wind speed forecasting, aims to predict future spatiotemporal data conditions based on historical spatiotemporal data observations, where the spatiotemporal data involved are detailed information characterizing wind speed. Currently, spatiotemporal forecasting technology has achieved many encouraging results. However, existing wind speed forecasting methods rarely consider the spatial dependencies between wind farms or have limited ability to mine spatiotemporal data dependencies, making it impossible to obtain satisfactory forecasting results. Summary of the invention

[0005] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] In view of the above existing problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by the present invention is that the existing wind speed prediction methods rarely consider the spatial dependency between wind farms or have limited ability to mine spatiotemporal data dependency, making it impossible to obtain satisfactory prediction results.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions: reading wind speed data and converting it into a matrix corresponding to three-dimensional wind speed data including time dimension, space dimension and feature dimension;

[0009] The three-dimensional matrix corresponding to the obtained wind speed data is processed by a first multi-layer perceptron to obtain first wind speed data;

[0010] The first wind speed data is processed by a K-layer spatiotemporal coding module to obtain wind speed data for a future time period to be predicted.

[0011] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, the matrix corresponding to the three-dimensional wind speed data can be expressed by the following formula:

[0012]

[0013] Among them, t is the current time, T1 is the historical time of wind speed data, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, and F is the feature dimension of the wind speed data.

[0014] As a preferred solution of the method for predicting wind speed data in a wind power generation scenario according to the present invention, the first wind speed data includes:

[0015]

[0016] in, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, n∈[1,N], and C is the feature dimension obtained after processing by the first multi-layer perceptron.

[0017] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, each layer in the K-layer spatiotemporal coding module processing includes at least two parallel branches, the first branch is a structured perception attention module, and the second branch is a frequency domain enhanced time delay perception attention module.

[0018] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, obtaining the wind speed data of the future time period to be predicted includes:

[0019] Normalize the wind speed data input to the Kth layer;

[0020] The wind speed data X input to the Kth layer k After mathematical model processing, the periodic information implicit in the wind speed data is obtained and trend informationX t k ;

[0021] The cycle information After being processed by the time delay perception module, a new periodic signal is obtained

[0022] The new periodic signal Perform attention processing in the frequency domain and obtain

[0023] Will and X t k Add together to get the wind speed signal

[0024] Through the serial operation of K time-space encoding modules, the final wind speed signal is obtained

[0025] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, the normalization process includes:

[0026]

[0027] Among them, GNN (g) is a graph neural network model, W Q , W K and W V is a learnable parameter, softmax(g) is a normalized exponential function, Trans(·) is a matrix transposition operation, and X o k is the normalized wind speed data X k .

[0028] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, the periodic information implicit in the wind speed data is obtained. and trend informationX t k ,include:

[0029] X t k =AvgPool(Padding(X k ))

[0030]

[0031] Among them, AvgPool(g) is a moving average operation, and Padding(g) is a matrix filling operation.

[0032] As a preferred solution of the wind speed data prediction method in the wind power generation scenario described in the present invention, the output of the k-th spatiotemporal coding module is obtained by using the following mathematical formula:

[0033]

[0034] Among them, MLP(g) is a multi-layer perceptron, Swish(g) and ReLU(g) are activation functions.

[0035] Beneficial effects of the present invention:

[0036] 1. By converting the one-dimensional wind speed data into a three-dimensional matrix, it not only includes the changes in the time series (time dimension), but also incorporates the information of spatial distribution (spatial dimension), as well as possible other characteristic dimensions (such as temperature, humidity, etc.). Such a conversion provides structured and multi-dimensional input for subsequent data processing and analysis, which helps to more comprehensively understand and predict wind speed changes;

[0037] 2. The three-dimensional wind speed data is processed by a multi-layer perceptron to achieve feature extraction and dimensionality reduction of the wind speed data. The multi-layer perceptron learns the inherent laws and patterns in the wind speed data through nonlinear transformation and weighted combination, thereby obtaining more abstract and refined first wind speed data, providing more effective input for subsequent predictions and improving the performance and accuracy of the prediction model;

[0038] 3. The first wind speed data is further processed through the K-layer spatiotemporal coding module to realize the spatiotemporal correlation modeling of the wind speed data. The spatiotemporal coding module takes into account the correlation of wind speed data in time and space. By learning and simulating these relationships, the prediction of wind speed data for the predicted future time period is obtained. It not only considers the trend and pattern of historical wind speed data, but also incorporates spatial correlation, making the prediction result more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0040] Figure 1 It is a flow chart of a method for predicting wind speed data in a wind power generation scenario shown in the present invention;

[0041] Figure 2 This is a schematic diagram comparing the effects of the method of the present invention and other classical methods. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0043] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without making any creative work should fall within the scope of protection of the present invention.

[0044] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0045] Example 1

[0046] According to an embodiment of the present invention, Figure 1 The flowchart shown is a method for predicting wind speed data in a wind power generation scenario, which specifically includes the following steps:

[0047] S1. Read the wind speed data and convert it into a matrix corresponding to the three-dimensional wind speed data including the time dimension, the space dimension and the feature dimension. It should be noted that the corresponding matrix can be expressed by the following formula:

[0048]

[0049] Among them, t is the current time, T1 is the historical time of wind speed data, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, and F is the feature dimension of the wind speed data.

[0050] As an example, three-dimensional wind speed data includes at least features including wind speed, wind direction, temperature, precipitation, air pressure, etc., acquired by N sensors at different times.

[0051] Preferably, the embodiment of the present invention converts the one-dimensional wind speed data into a three-dimensional matrix, which not only includes the changes in the time series (time dimension), but also incorporates the information of spatial distribution (spatial dimension), as well as possible other characteristic dimensions (such as temperature, humidity, etc.). Such conversion provides structured and multi-dimensional input for subsequent data processing and analysis, which helps to more comprehensively understand and predict wind speed changes.

[0052] S2. Process the three-dimensional matrix corresponding to the obtained wind speed data through a first multi-layer perceptron to obtain first wind speed data. It should be noted that in this step, the first wind speed data can be represented by the following formula:

[0053]

[0054] in, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, n∈[1,N], and C is the feature dimension obtained after processing by the first multi-layer perceptron.

[0055] Preferably, this step processes the three-dimensional wind speed data through a multi-layer perceptron, thereby realizing feature extraction and dimensionality reduction of the wind speed data. The multi-layer perceptron learns the inherent laws and patterns in the wind speed data through nonlinear transformation and weighted combination, thereby obtaining a more abstract and refined first wind speed data, providing a more effective input for subsequent predictions, and improving the performance and accuracy of the prediction model.

[0056] S3, the first wind speed data is processed by a K-layer spatiotemporal coding module to obtain the wind speed data of the future time period to be predicted. It should be noted that this step obtains the wind speed data of the future time period to be predicted, including:

[0057] The wind speed data input to the Kth layer is normalized, and its mathematical expression formula is as follows:

[0058]

[0059] Among them, GNN (g) is a graph neural network model, W Q , W K and W V is a learnable parameter, softmax(g) is a normalized exponential function, Trans(·) is a matrix transposition operation, and X o k is the normalized wind speed data X k ;

[0060] The wind speed data X input to the Kth layer k After mathematical model processing, the periodic information implicit in the wind speed data is obtained and trend informationX t k , its mathematical expression formula is as follows:

[0061] X t k =AvgPool(Padding(X k ))

[0062]

[0063] Among them, AvgPool(g) is a moving average operation, and Padding(g) is a matrix filling operation;

[0064] Cycle Information After being processed by the time delay perception module, a new periodic signal is obtained in:

[0065] (1) The historical wind speed data are divided into a group of representative short-term wind speed patterns. For example, the historical wind speed data are sliced ​​using a sliding window of size S to obtain a group of wind speed flow sequences, which are expressed as follows: The length of each segment is S,N p is the number of wind speed data slices;

[0066] (2) The k-Shape clustering algorithm is used to divide the obtained wind speed flow sequence into N p Class, that is Among them, p i It is also a time series of length S, i∈[1,N p ] is the i-th category;

[0067] (3) Based on the above N p time series and N p Classes are integrated using the following formula:

[0068]

[0069] Among them, W t1 , W t2 , W1, W2, is a learnable parameter;

[0070] (4) The periodic signal X t k Adding it to R gives a new periodic signal representation

[0071] The new periodic signal Perform attention processing in the frequency domain and obtain Its mathematical expression formula is as follows:

[0072]

[0073] Among them, softmax(g) is the normalized exponential function, F(g) is the fast Fourier transform, and F -1 (g) is the fast inverse Fourier transform, W FQ , W FK and W FK is a learnable matrix;

[0074] Will and X t k Add together to get the wind speed signal

[0075] The output of the k-th spatiotemporal coding module is obtained using the following mathematical formula:

[0076]

[0077] Among them, MLP(g) is a multi-layer perceptron, Swish(g) and ReLU(g) are activation functions;

[0078] Through the serial operation of K time-space encoding modules, the final wind speed signal is obtained The specific expression is as follows:

[0079]

[0080] in, are the outputs of 1, 2, …, K space-time coding modules respectively.

[0081] As an example, each layer in the K-layer spatiotemporal coding module processing includes at least two parallel branches, the first branch is a structured perception attention module, and the second branch is a frequency domain enhanced time delay perception attention module.

[0082] In order to verify the superiority of the wind speed data prediction method in a wind power generation scenario proposed in an embodiment of the present invention, this embodiment uses the invention and the classic differential autoregressive moving average model (ARIMA), support vector regression model (SVR), long short-term memory network (LSTM), attention spatiotemporal graph neural network model (ASTGCN), and graph wavelet model (GW) to test the spatiotemporal data sets PEMS04, PEMS07, and PEMS08, and uses MAE, MAPE, and RMSE to evaluate their effects.

[0083] Reference Figure 2 It can be directly seen that ARIMA requires the spatiotemporal data used for prediction to meet specific linear assumptions, but it is well known that spatiotemporal data are too complex and generally cannot meet these linear assumptions. Therefore, they usually perform poorly in practice; traditional machine learning models represented by SVM are used to process complex nonlinear spatiotemporal data due to their nonlinear processing capabilities, thereby obtaining better spatiotemporal prediction effects. However, such algorithms are extremely dependent on the choice of kernel functions; the LSTM model regards the spatiotemporal network as a grid-type structure similar to an image, which obviously violates the actual structure of the spatiotemporal network, so the wind speed prediction effect is not good; the ASTGCN and GW methods are wind speed prediction models based on graph neural networks, which still have great deficiencies in mining spatiotemporal dependencies.

[0084] The wind speed data prediction method disclosed in the embodiment of the present invention adopts a structured perception attention module to mine the spatial dependency in the spatiotemporal sequence, and at the same time, adopts a frequency domain enhanced time delay perception attention module to mine the temporal dependency in the spatiotemporal sequence.

[0085] pass Figure 2It can be seen that for the PEMS04 data set, the MAE, RMSE and MAPE of the ARIMA algorithm are 23.28, 37.39 and 19.79% respectively, while the MAE, RMSE and MAPE of the SVM algorithm are 22.27, 37.43 and 19.09% respectively; for the PEMS07 data set, the MAE, RMSE and MAPE of the ARIMA algorithm are 34.11, 51.46 and 19.68% respectively, while the MAE, RMSE and MAPE of the SVM algorithm are 30 .90, 45.74 and 16.69%; for the PEMS08 dataset, the MAE, RMSE and MAPE of the ARIMA algorithm are 22.36, 34.52 and 14.47% respectively, while the MAE, RMSE and MAPE of the SVM algorithm are 22.26, 34.50 and 14.38% respectively; From the above data, it can be seen that the two algorithms perform the worst among all the compared algorithms. The main reason is that their inherent linearity or strict data requirements make them unable to handle complex spatiotemporal data well.

[0086] LSTM is a typical deep learning model. For the PEMS04 dataset, its MAE, RMSE and MAPE are 22.03, 36.90 and 18.21% respectively; for the PEMS07 dataset, its MAE, RMSE and MAPE are 24.38, 39.76 and 11.16% respectively; for PEMS08, its MAE, RMSE and MAPE are 23.19, 33.27 and 13.69% respectively. From the above data, it can be seen that although its performance is better than that of ARIMA and SVM algorithms, it regards the complex spatiotemporal series network as a grid data, which makes its performance worse than that of the model based on graph neural network.

[0087] In addition, ASTGCN and GW are spatiotemporal prediction models based on graph neural networks. For the PEMS04 dataset, the MAE, RMSE and MAPE of the ASTGCN model are 21.89, 34.31 and 16.13%, respectively, and the MAE, RMSE and MAPE of the GW model are 19.17, 31.95 and 13.28%, respectively; for the PEMS07 dataset, the MAE, RMSE and MAPE of the ASTGCN algorithm are 22.49, 35.76 and 9.89%, respectively, while the MAE, RMSE and MAPE of the GW algorithm are The APEs are 30.90, 45.74 and 16.69% respectively; for the PEMS08 dataset, the MAE, RMSE and MAPE of the ASTGCN algorithm are 19.75, 29.36 and 12.89% respectively, while the MAE, RMSE and MAPE of the GW algorithm are 15.97, 24.81 and 10.27% respectively; From the above data, it can be seen that the algorithm based on graph neural network has certain advantages over the previous algorithms; however, due to the poor acquisition of complex spatiotemporal dependencies in spatiotemporal data by such algorithms, the performance still needs to be improved.

[0088] Preferably, the model proposed in the embodiment of the present invention has MAE, RMSE and MAPE of 18.74, 30.92 and 12.37% for the PEMS04 dataset, MAE, RMSE and MAPE of 19.75, 33.43 and 8.21% for the PEMS07 dataset, and MAE, RMSE and MAPE of 14.96, 24.17 and 9.83% for the PEMS08 dataset. In summary, compared with other comparison algorithms, the SAGSFEAT model provided in the embodiment of the present invention has better spatiotemporal prediction performance.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting wind speed data in a wind power generation scenario, It is characterized in that include: Read wind speed data and convert it into a matrix corresponding to three-dimensional wind speed data including time dimension, space dimension and feature dimension; The three-dimensional matrix corresponding to the obtained wind speed data is processed by a first multi-layer perceptron to obtain first wind speed data; The first wind speed data is processed by a K-layer spatiotemporal coding module to obtain wind speed data for a future time period to be predicted.

2. The method for predicting wind speed data in a wind power generation scenario according to claim 1, It is characterized in that The matrix corresponding to the three-dimensional wind speed data can be expressed by the following formula: X′=(X′ t-T1+1 ,X′ t-T1+2 ,…,X′ t ) Where t is the current time, T 1 is the historical time of wind speed data, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, and F is the feature dimension of the wind speed data.

3. The method for predicting wind speed data in a wind power generation scenario according to claim 1, It is characterized in that The first wind speed data includes: in, represents the wind speed characteristics captured by N sensors at time t, represents the wind speed feature captured by the nth sensor at time t, n∈[1,N], and C is the feature dimension obtained after processing by the first multi-layer perceptron.

4. The method for predicting wind speed data in a wind power generation scenario according to claim 1, It is characterized in that Each layer in the K-layer spatiotemporal coding module processing includes at least two parallel branches, the first branch is a structured perception attention module, and the second branch is a frequency domain enhanced time delay perception attention module.

5. The method for predicting wind speed data in a wind power generation scenario according to claim 1, It is characterized in that Get the wind speed data for the future time period to be predicted, including: Normalize the wind speed data input to the Kth layer; The wind speed data X input to the Kth layer k After mathematical model processing, the periodic information implicit in the wind speed data is obtained and trend informationX t k ; The cycle information After being processed by the time delay perception module, a new periodic signal is obtained The new periodic signal Perform attention processing in the frequency domain and obtain Will and Add together to get the wind speed signal Through the serial operation of K time-space encoding modules, the final wind speed signal is obtained 6. The method for predicting wind speed data in a wind power generation scenario according to claim 5, It is characterized in that The normalization process includes: Among them, GNN (g) is a graph neural network model, W Q , W K and W V is a learnable parameter, softmax(g) is a normalized exponential function, Trans(·) is a matrix transposition operation, and X o k is the normalized wind speed data X k .

7. The method for predicting wind speed data in a wind power generation scenario according to claim 5, It is characterized in that Get the period information implicit in the wind speed data and trend informationX t k ,include: X t k =AvgPool(Padding(X k )) Among them, AvgPool(g) is a moving average operation, and Padding(g) is a matrix filling operation.

8. The method for predicting wind speed data in a wind power generation scenario according to claim 5, It is characterized in that The output of the k-th spatiotemporal coding module is obtained using the following mathematical formula: Among them, MLP(g) is a multi-layer perceptron, Swish(g) and ReLU(g) are activation functions.