Adjacent wind speed prediction method for strong convection wind

Abstract

The invention particularly relates to an adjacent wind speed prediction method for strong convection strong wind, and belongs to the technical field of wind speed prediction. According to the method, near wind field forecasting is realized based on the improved Transform model, and the model can better fuse information among multi-source data and learn change characteristics of the data in a spatial scale and a time scale by constructing modules such as the MSAB and the FGU, so that the nowcasting effect of the final wind speed is improved. Compared with CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network) deep learning models, the deep learning model used by the method can effectively extract spatial distribution characteristics of data, and can also be suitable for long-term sequence prediction tasks. And meanwhile, based on a transformer architecture, matrix budget is used for replacing circulation in a traditional RNN, so that the model can be subjected to parallel calculation, and the calculation efficiency is improved.

Description

Technical Field

[0001] The invention specifically relates to a method for predicting the near-surface wind speed of a strong convective gale, and belongs to the technical field of wind speed prediction. Background Art

[0002] Severe convective wind is one of the main meteorological disasters that cause power system disasters. In the vicinity of dense transmission channels, due to the dense transmission lines, it is more likely to have wind deflection and tower collapse in the face of extreme strong winds, thus affecting the stability of the power supply system and causing losses to the power sector. Therefore, according to the characteristics of strong convective winds, it is effective to predict the wind speed of the wind (wind speed in the next 3 hours), which is of great significance to help people take corresponding preventive measures and reduce the loss of life and property of the people. Most of the existing technologies use deep learning models for wind speed prediction, mainly based on two types of models: convolutional neural network and recurrent neural network. Although the convolutional neural network model can fully extract the spatial distribution characteristics of sequence data by superimposing convolution modules and has good adaptability to the spatial size of data, it is difficult to learn and associate the change law of the sequence in the time dimension and store it in memory, so it is not suitable for long-term sequence prediction tasks. Although the recurrent neural network can learn the spatiotemporal changes of sequence data through gating mechanisms and recursive algorithms, it is limited by the local receptive field of the current time and the feature information in the storage unit. It is difficult to grasp the key change characteristics of the overall data in the sequence and learn effective long-term dependencies, and its performance is limited in the application of long-term sequence prediction. Based on the above shortcomings, the existing deep learning models still have the problem of insufficient prediction accuracy when used for the near-term wind speed prediction of severe convective winds. Summary of the invention

[0003] The technical problem to be solved by the present invention is: how to improve the accuracy of the prediction of the approaching wind speed of strong convective gales.

[0004] In order to solve the above technical problems, the technical solution proposed by the present invention is: a method for predicting the approaching wind speed of a strong convective gale, comprising the following steps:

[0005] Step 1: Establish the initial near-wind speed prediction model based on the Transformer neural network and multi-scale spatiotemporal attention algorithm, as shown in the following formula (1):

[0006] (1)

[0007] In formula (1), X and Y are the radar reflectivity and wind speed data to be input respectively; It is the radar reflectivity code with the position information superimposed. It is the wind speed code after superimposing the position information; MSAB ( ) is a multi-scale spatiotemporal attention module, ( ) is the Fourier fusion module; It is the intermediate information of radar reflectivity encoding after two attention mechanisms. It is the intermediate information of wind speed encoding after two attention mechanisms; Z is the radar reflectivity and wind speed fusion information after fusion by the Fourier gating unit; O is the predicted wind speed output by the initial adjacent wind speed prediction model;

[0008] Step 2: Continuously collect data N times, and pre-process the collected N historical wind speed data and the N historical radar reflectivity data corresponding to the N historical wind speed data respectively; collect the pre-processed N historical radar reflectivity data to form a historical radar reflectivity data set X, as shown in the following formula (2):

[0009] (2)

[0010] In formula (2), They are the first historical radar reflectivity, the second historical radar reflectivity to the Nth historical radar reflectivity collected and preprocessed;

[0011] The N pre-processed historical wind speed data are collected to form the historical wind speed data set Y, as shown in the following formula (3):

[0012] (3)

[0013] In formula (3), They are the first historical wind speed, the second historical wind speed to the Nth historical wind speed collected and preprocessed;

[0014] Step 3: Divide the historical radar reflectivity data set X into a first training set, a first test set and a first validation set in sequence according to an empirical ratio; divide all data in the historical wind speed data set Y into a second training set, a second test set and a second validation set in sequence according to the same empirical ratio; substitute the first training set and the second training set into the initial near-by wind speed prediction model for model training; substitute the first test set and the second test set into the initial near-by wind speed prediction model for model testing; substitute the first validation set and the second validation set into the initial near-by wind speed prediction model for model verification, and obtain a final near-by wind speed prediction model that has been trained; the final near-by wind speed prediction model that has been trained uses the historical radar reflectivity data and the historical wind speed data before the period of wind speed prediction for the severe convective gale to be performed as input, and uses the wind speed of the period of wind speed prediction for the severe convective gale to be performed as output;

[0015] Step 4: Collect historical radar reflectivity data and historical wind speed data at m moments before the period of wind speed prediction for the severe convective gale in real time; substitute the m historical radar reflectivity data and historical wind speed data into the trained final near-term wind speed prediction model to obtain the wind speed for the period of fractional prediction for the severe convective gale.

[0016] Furthermore, the multi-scale spatiotemporal attention module in the initial near wind speed prediction model in step 1 is as shown in the following formula (4):

[0017] (4)

[0018] In formula (4), is the output of the multi-scale spatiotemporal attention module; It is a multi-layer perceptron; is a fully connected layer; is an intermediate variable containing time information; It is an intermediate variable containing global space information; It is the temporal attention mechanism; It is a sliding window attention mechanism; It is the window attention mechanism.

[0019] Furthermore, the Fourier fusion module in the initial approaching wind speed prediction model in step 1 is as shown in the following formula (5):

[0020] (5)

[0021] In formula (5), is the output of the Fourier fusion module; is a regularization method; H is (feature information); k is a gating unit used to adjust the update information flow during data fusion; is the sigmoid activation function; , and They are the first convolution, the second convolution, and the third convolution respectively; It is the merging of data in the channel dimension; q is the gating unit, which is used to adjust the fusion information output by AFNO and the detail information output by the large convolution; It is a depth-wise separable large convolution; is the Fourier neural operator; It is the superposition information of wind speed and radar; It is the output information after being processed by Fourier neural operator; It is the fully connected layer acting on the radar; is a fully connected layer acting on wind speed.

[0022] Further, the preprocessing process in step 2 includes the following contents:

[0023] 1) Based on the visualization image of radar reflectivity, the clutter data is eliminated on the basis of considering the continuity of data, and the radar echo data containing weather processes is retained;

[0024] 2) Gridding the discrete wind speed data, interpolating the data onto the grid using the inverse distance weighted interpolation method, and matching it with the radar reflectivity data in time and space;

[0025] 3) Normalization: The historical wind speed data and historical wind field observation characteristic data are normalized through maximum and minimum normalization, and the data are scaled to the range of [0,1].

[0026] Further, the training process of the initial approaching wind speed prediction model in step 3 is as follows:

[0027] Step 3.1: input the first training set and the second training set into the initial near-wind speed prediction model for batch training, divide the data in the first training set and the second training set into multiple batches again, and update the parameters of the model on each batch; during the training process of each batch, the model calculates the loss function according to the current input data and parameters, and updates the parameters of the model according to the optimizer;

[0028] The loss function is shown in the following formula (6):

[0029] (6)

[0030] In formula (6), is the total loss function, Calculate the weighted root mean square error between the predicted echo sequence and the true value, W is the weight under different wind speed thresholds, is the predicted wind speed of the initial near-by wind speed prediction model, is the actual observed wind speed; is the generator loss, which is obtained by sending the predicted echo sequence and the true value sequence into the discriminator for discrimination. ; is the perceptual similarity loss, where and are the encoding vectors of the predicted echo sequence and the true value, respectively, which are learnable. The root mean square error between the two is calculated to obtain ; and is the regularization coefficient;

[0031] Step: 3.2: The model needs to be trained for multiple rounds to be fully fitted. After each round of training, the effect is tested using the data from the first validation set and the second validation set. Models with qualified prediction effects will be saved for wind speed prediction, otherwise iterative training will continue;

[0032] Step 3.3: Use the final saved model and the first test set and the second test set to further test the robustness of the model. The model with qualified prediction effect on the test machine will be used as the final near-wind speed prediction model after training. Otherwise, the model's learning rate, downsampling and other hyperparameters will be further modified, and the model will be retrained until the expected effect is achieved.

[0033] Beneficial effects of the present invention: 1. The present invention realizes the near-term wind farm forecast based on the improved Transformer model. By constructing modules such as MSAB and FGU, the model can better integrate information between multi-source data and learn the changing characteristics of data in spatial and temporal scales, thereby improving the near-term forecast effect of the final wind speed. 2. Compared with the two types of deep learning models, CNN and RNN, the deep learning model used in the present invention can not only effectively extract the spatial distribution characteristics of data, but also be applicable to long-term sequence prediction tasks. At the same time, based on the transformer architecture, matrix budget is used instead of the loop in the traditional RNN, so that the model can be calculated in parallel, thereby improving computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the method for predicting the approaching wind speed of severe convective gale of the present invention.

[0035] Figure 2 Schematic diagram of the network structure of the initial approaching wind speed prediction model in the first embodiment.

[0036] Figure 3 Schematic diagram of the network structure of the multi-scale spatiotemporal attention module in Example 1.

[0037] Figure 4 Schematic diagram of the network structure of the Fourier fusion module in the first embodiment. DETAILED DESCRIPTION

[0038] The following is a further description of a method for predicting the wind speed of a strong convective gale according to the present invention in conjunction with specific embodiments and accompanying drawings.

[0039] Example

[0040] The method for predicting the approaching wind speed of a strong convective gale in this embodiment is as follows: Figure 1 As shown, the following steps are included:

[0041] Step 1: Establish the initial wind speed prediction model based on the Transformer neural network and multi-scale spatiotemporal attention algorithm, as shown in the following formula (1). The network structure is as follows: Figure 2 As shown,

[0042] (1)

[0043] In formula (1), X and Y are the radar reflectivity and wind speed data to be input respectively; It is the radar reflectivity code with the position information superimposed. It is the wind speed code after superimposing the position information; MSAB ( ) is a multi-scale spatiotemporal attention module, ( ) is the Fourier fusion module; It is the intermediate information of radar reflectivity encoding after two attention mechanisms. It is the intermediate information of wind speed encoding after two attention mechanisms; Z is the radar reflectivity and wind speed fusion information after fusion by the Fourier gating unit; O is the predicted wind speed output by the initial adjacent wind speed prediction model;

[0044] The multi-scale spatiotemporal attention module in the initial near-wind speed prediction model is as follows (4), and the network structure is as follows Figure 3 As shown,

[0045] (4)

[0046] In formula (4), is the output of the multi-scale spatiotemporal attention module; It is a multi-layer perceptron; is a fully connected layer; is an intermediate variable containing time information; It is an intermediate variable containing global space information; It is the temporal attention mechanism; It is a sliding window attention mechanism; It is the window attention mechanism.

[0047] The Fourier fusion module in the initial approaching wind speed prediction model is as follows (5), and the network structure is as follows Figure 4 As shown,

[0048] (5)

[0049] In formula (5), is the output of the Fourier fusion module; is a regularization method; H is (feature information); k is a gating unit used to adjust the update information flow during data fusion; is the sigmoid activation function; , and They are the first convolution, the second convolution, and the third convolution respectively; It is the merging of data in the channel dimension; q is the gating unit, which is used to adjust the fusion information output by AFNO and the detail information output by the large convolution; It is a depth-wise separable large convolution; is the Fourier neural operator; It is the superposition information of wind speed and radar; It is the output information after being processed by Fourier neural operator; It is the fully connected layer acting on the radar; is a fully connected layer acting on wind speed.

[0050] Step 2: Continuously collect data N times, and pre-process the collected N historical wind speed data and the N historical radar reflectivity data corresponding to the N historical wind speed data respectively;

[0051] The preprocessing process includes the following:

[0052] 1) Based on the visualization image of radar reflectivity, the clutter data is eliminated on the basis of considering the continuity of data, and the radar echo data containing weather processes is retained;

[0053] 2) Gridding the discrete wind speed data, interpolating the data onto the grid using the inverse distance weighted interpolation method, and matching it with the radar reflectivity data in time and space;

[0054] 3) Normalization: The historical wind speed data and historical wind field observation characteristic data are normalized through maximum and minimum normalization, and the data are scaled to the range of [0,1].

[0055] The N pre-processed historical radar reflectivity data are collected to form a historical radar reflectivity data set X, as shown in the following formula (2):

[0056] (2)

[0057] In formula (2), They are the first historical radar reflectivity, the second historical radar reflectivity to the Nth historical radar reflectivity collected and preprocessed;

[0058] The N pre-processed historical wind speed data are collected to form the historical wind speed data set Y, as shown in the following formula (3):

[0059] (3)

[0060] In formula (3), They are the first historical wind speed, the second historical wind speed to the Nth historical wind speed collected and preprocessed;

[0061] Step 3: Divide the historical radar reflectivity data set X into a first training set, a first test set and a first validation set according to the empirical ratio; divide all the data in the historical wind speed data set Y into a second training set, a second test set and a second validation set according to the same empirical ratio; substitute the first training set and the second training set into the initial near wind speed prediction model for model training; substitute the first test set and the second test set into the initial near wind speed prediction model for model testing; substitute the first validation set and the second validation set into the initial near wind speed prediction model for model verification, and obtain the final near wind speed prediction model after training;

[0062] The training process of the initial near-wind speed prediction model is as follows:

[0063] Step 3.1: Input the first training set and the second training set into the initial near-wind speed prediction model for batch training, divide the data in the first training set and the second training set into multiple batches again, and update the parameters of the model on each batch; during the training process of each batch, the model calculates the loss function according to the current input data and parameters, and updates the parameters of the model according to the optimizer;

[0064] The loss function is as follows (6):

[0065] (6)

[0066] In formula (6), is the total loss function, Calculate the weighted root mean square error between the predicted echo sequence and the true value, W is the weight under different wind speed thresholds, is the predicted wind speed of the initial near wind speed prediction model, is the actual observed wind speed; is the generator loss, which is obtained by sending the predicted echo sequence and the true value sequence into the discriminator for discrimination. ; is the perceptual similarity loss, where and are the encoding vectors of the predicted echo sequence and the true value, respectively, which are learnable. The root mean square error between the two is calculated to obtain ; and is the regularization coefficient;

[0067] Step: 3.2: The model needs to be trained for multiple rounds to be fully fitted. After each round of training, the effect is tested using the data from the first validation set and the second validation set. Models with qualified prediction effects will be saved for wind speed prediction, otherwise iterative training will continue;

[0068] Step 3.3: Use the final saved model and the first test set and the second test set to further test the robustness of the model. The model with qualified prediction effect on the test machine will be used as the final near-wind speed prediction model after training. Otherwise, the model's learning rate, downsampling and other hyperparameters will be further modified, and the model will be retrained until the expected effect is achieved.

[0069] The final near-term wind speed prediction model after training takes the historical radar reflectivity data and historical wind speed data before the period of strong convective wind to be predicted as input, and takes the wind speed during the period of strong convective wind to be predicted as output;

[0070] Step 4: Collect historical radar reflectivity data and historical wind speed data at m moments before the period of strong convective winds to be predicted in real time; substitute the m historical radar reflectivity data and historical wind speed data into the trained final near-term wind speed prediction model to obtain the wind speed in the period of strong convective winds to be predicted by scores.

Claims

1. A method for predicting the approaching wind speed of a severe convective gale, characterized in that: The following steps are involved: Step 1: Establish the initial approaching wind speed prediction model based on the Transformer neural network and multi-scale spatiotemporal attention algorithm, as shown in the following formula (1): （1） In formula (1), X and Y are the radar reflectivity and wind speed data to be input respectively; It is the radar reflectivity code with the position information superimposed. It is the wind speed code after superimposing the position information; MSAB ( ) is a multi-scale spatiotemporal attention module, ( ) is the Fourier fusion module; It is the intermediate information of radar reflectivity encoding after two attention mechanisms. It is the intermediate information of wind speed encoding after two attention mechanisms; Z is the radar reflectivity and wind speed fusion information after fusion by the Fourier gating unit; O is the predicted wind speed output by the initial adjacent wind speed prediction model; Step 2: Continuously collect data N times, and pre-process the collected N historical wind speed data and the N historical radar reflectivity data corresponding to the N historical wind speed data respectively; collect the pre-processed N historical radar reflectivity data to form a historical radar reflectivity data set X, as shown in the following formula (2): （2） In formula (2), They are the first historical radar reflectivity, the second historical radar reflectivity to the Nth historical radar reflectivity collected and preprocessed; The N pre-processed historical wind speed data are collected to form the historical wind speed data set Y, as shown in the following formula (3): （3） In formula (3), They are the first historical wind speed, the second historical wind speed to the Nth historical wind speed collected and preprocessed; Step 3: Divide the historical radar reflectivity data set X into a first training set, a first test set and a first validation set in sequence according to an empirical ratio; divide all data in the historical wind speed data set Y into a second training set, a second test set and a second validation set in sequence according to the same empirical ratio; substitute the first training set and the second training set into the initial near-by wind speed prediction model for model training; substitute the first test set and the second test set into the initial near-by wind speed prediction model for model testing; substitute the first validation set and the second validation set into the initial near-by wind speed prediction model for model verification, and obtain a final near-by wind speed prediction model that has been trained; the final near-by wind speed prediction model that has been trained uses the historical radar reflectivity data and the historical wind speed data before the period of wind speed prediction for the severe convective gale to be performed as input, and uses the wind speed of the period of wind speed prediction for the severe convective gale to be performed as output; Step 4: Collect historical radar reflectivity data and historical wind speed data at m moments before the period of wind speed prediction for the severe convective gale in real time; substitute the m historical radar reflectivity data and historical wind speed data into the trained final near-term wind speed prediction model to obtain the wind speed for the period of fractional prediction for the severe convective gale.

2. The near-term wind speed prediction method according to claim 1, characterized in that: The multi-scale spatiotemporal attention module in the initial near wind speed prediction model in step 1 is shown in the following formula (4): （4） In formula (4), is the output of the multi-scale spatiotemporal attention module; It is a multi-layer perceptron; is a fully connected layer; is an intermediate variable containing time information; It is an intermediate variable containing global space information; It is the temporal attention mechanism; It is a sliding window attention mechanism; It is the window attention mechanism.

3. The near-term wind speed prediction method according to claim 1, characterized in that: The Fourier fusion module in the initial approaching wind speed prediction model in step 1 is shown in the following formula (5): （5） In formula (5), is the output of the Fourier fusion module; is a regularization method; H is (feature information); k is a gating unit used to adjust the update information flow during data fusion; is the sigmoid activation function; , and They are the first convolution, the second convolution, and the third convolution respectively; It is the merging of data in the channel dimension; q is the gating unit, which is used to adjust the fusion information output by AFNO and the detail information output by the large convolution; It is a depth-wise separable large convolution; is the Fourier neural operator; It is the superposition information of wind speed and radar; It is the output information after being processed by Fourier neural operator; It is the fully connected layer acting on the radar; is a fully connected layer acting on wind speed.

4. The near-term wind speed prediction method according to claim 1, characterized in that: The preprocessing process in step 2 includes the following contents: 1) Based on the visualization image of radar reflectivity, the clutter data is eliminated on the basis of considering the continuity of data, and the radar echo data containing weather processes is retained; 2) Gridding the discrete wind speed data, interpolating the data onto the grid using the inverse distance weighted interpolation method, and matching it with the radar reflectivity data in time and space; 3) Normalization: The historical wind speed data and historical wind field observation characteristic data are normalized through maximum and minimum normalization, and the data are scaled to the range of [0,1].

5. The near-term wind speed prediction method according to claim 1, characterized in that: The training process of the initial near-term wind speed prediction model in step 3 is as follows: Step 3.1: input the first training set and the second training set into the initial near-wind speed prediction model for batch training, divide the data in the first training set and the second training set into multiple batches again, and update the parameters of the model on each batch; during the training process of each batch, the model calculates the loss function according to the current input data and parameters, and updates the parameters of the model according to the optimizer; The loss function is shown in the following formula (6): （6） In formula (6), is the total loss function, Calculate the weighted root mean square error between the predicted echo sequence and the true value, W is the weight under different wind speed thresholds, is the predicted wind speed of the initial near-by wind speed prediction model, is the actual observed wind speed; is the generator loss, which is obtained by sending the predicted echo sequence and the true value sequence into the discriminator for discrimination. ; is the perceptual similarity loss, where and are the encoding vectors of the predicted echo sequence and the true value, respectively, which are learnable. The root mean square error between the two is calculated to obtain ; and is the regularization coefficient; Step: 3.2: The model needs to be trained for multiple rounds to be fully fitted. After each round of training, the effect is tested using the data from the first validation set and the second validation set. Models with qualified prediction effects will be saved for wind speed prediction, otherwise iterative training will continue; Step 3.3: Use the final saved model and the first test set and the second test set to further test the robustness of the model. The model with qualified prediction effect on the test machine will be used as the final near-wind speed prediction model after training. Otherwise, the model's learning rate, downsampling and other hyperparameters will be further modified, and the model will be retrained until the expected effect is achieved.

Images