Wind speed prediction method based on numerical weather forecast

Abstract

A wind speed prediction method based on numerical weather forecast relates to the technical field of wind speed prediction, and comprises the following steps: firstly, collecting NWP data of a plurality of grid nodes in a target wind field and operation data of fans in the wind field in the same time period, and establishing feature vectors; constructing a wind speed prediction model; the feature vectors are preprocessed and then converted into feature images; and finally inputting the feature image into the wind speed prediction model to obtain a wind speed prediction result. The method can improve the precision of wind speed prediction.

Description

Technical Field

[0001] The invention relates to the technical field of wind speed prediction, and in particular to a wind speed prediction method based on numerical weather forecasting. Background Art

[0002] As the wind energy industry continues to advance, the cost of wind power generation is gradually decreasing. The construction of new power systems has put forward higher accuracy requirements for wind power forecasts in wind farms. Accurate wind power forecasts rely on accurate predictions of wind speed data.

[0003] Wind speed prediction is affected by numerical weather prediction (NWP) data, but the resolution of NWP is low at present. In addition, affected by topography, the wind speeds of different units in the same wind farm vary greatly, resulting in local information loss and unable to meet the requirements of micro-scale wind speed prediction. In addition, the single time series prediction model is trained with historical wind speed series as input, contains less information, lacks meteorological background and spatial information, and has poor prediction effect. Therefore, by fusing NWP and wind turbine wind speed time series information, it contains more spatiotemporal information and can more accurately predict the wind speed of the unit.

[0004] At present, scholars at home and abroad have conducted a lot of research on wind speed prediction and proposed various methods to improve the accuracy of wind speed prediction. Wind speed prediction methods can be roughly divided into three categories: physical methods, statistical methods, and machine learning and deep learning methods.

[0005] Among them, the physical method is based on NWP and regional geographic information, using physical laws to simulate atmospheric motion and predict wind speed. However, due to the resolution limitation of NWP, the physical method is mainly effective in forecasting on a large spatial scale. In addition, the physical method needs to solve a large number of atmospheric motion equations and consider additional background information, so it often requires a lot of computing resources.

[0006] The statistical method uses certain historical data (wind turbine operation data, meteorological data, etc.) to train the mapping function of wind speed prediction. Traditional statistical methods mainly include autoregressive model (AR), autoregressive moving average model (ARMA), autoregressive intergrated moving average model (ARIMA), etc. However, for exponential data or fluctuating data, the difference operation adopted by the ARIMA model will lead to information loss and reduced fitting effect.

[0007] With the development of artificial intelligence, machine learning and deep learning methods have been widely used in wind speed prediction. For example, back propagation neural network (BPNN), support vector machine (SVM), extreme gradient boosting regression tree algorithm (eXtreme Gradient Boosting, XGBoost), etc. Considering that the advantages and disadvantages of each machine learning algorithm are relatively obvious, and there are certain applicable scenarios and scopes, more studies choose to adopt the method of combining prediction models to complement each other's advantages and improve prediction accuracy. In addition, some researchers have tried to combine decision trees with k-nearest neighbor algorithms and compared them with the SVR model with the best single effect. The final result was improved by 37%, proving that the effect of simple model combination prediction is far better than that of a single machine learning model. Therefore, compared with traditional statistical models, models based on deep learning show better prediction performance.

[0008] The literature [Liu H, Mi X, Li Y. Smart deep learning based wind speed prediction model using wavelet packet decomposition, convolutional neural network and convolutional long short term memory network [J]. Energy Conversion and Management, 2018, 166: 120-131] proposed a prediction model combining LSTM and CNN to obtain the wind speed of different frequency sublayers through a data decomposition algorithm.

[0009] The literature [Zhang Yanni, Shi Jiarong, Li Jin, et al. Short-term wind speed prediction by integrating residuals with VMD-ELM-LSTM [J]. Acta Energiae Solaris Sinica, 2023, 44(09): 340-347] integrates residuals with variational mode decomposition (VMD), extreme learning machine (ELM), and long short-term memory (LSTM) to improve the accuracy of wind speed prediction.

[0010] The literature [Zhang YM, Wang H. Multi-head attention-based probabilistic CNN-BiLSTM for day-ahead wind speed forecasting [J]. Energy, 2023, 278: 127865.] proposed a probabilistic model based on CNN-BiLSTM and deep integration strategy to predict the day-ahead wind speed. Compared with 9 probabilistic benchmark methods, it has better prediction accuracy and uncertainty estimation ability.

[0011] Reference [Wang Shiming, Zhang Shaotong, Lou Jiayi. Research on ultra-short-term wind speed prediction based on CNN-LSTM-ARIMA [J]. New Energy Progress: 1-9] uses ARIMA analysis of historical residuals to correct the prediction value of the CNN-LSTM model, which improves the accuracy of the model. However, the applicability of the above method to wind speed prediction of micro-scale units still needs to be improved.

[0012] A Chinese patent (CN114358398A) discloses a method for correcting wind speed in numerical weather forecasts based on a deep neural network. Through dynamic correction, the problem of insufficient data feature extraction caused by a small amount of data is compensated, and the error of the NWP wind speed prediction value is reduced, thereby reducing the impact of the error of the NWP wind speed data on the wind power prediction results. The method proposed in the present invention has strong versatility and can effectively correct a large number of NWP wind speed prediction values; however, the machine learning model in this method is relatively simple and cannot accurately predict the wind speed;

[0013] Therefore, we propose a method to improve the wind speed prediction accuracy. Summary of the invention

[0014] In order to overcome the deficiencies in the background technology, the present invention discloses a wind speed prediction method based on numerical weather forecasting.

[0015] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme:

[0016] A wind speed prediction method based on numerical weather forecast, the specific steps are:

[0017] Collect NWP data of several grid nodes in the target wind farm, as well as the operation data of wind turbines in the wind farm during the same period, and establish feature vectors;

[0018] Construct wind speed prediction model;

[0019] The feature vector is converted into a feature image after preprocessing;

[0020] Input the feature image into the wind speed prediction model to obtain the wind speed prediction result.

[0021] Preferably, the preprocessing process includes outlier processing, missing value processing and data normalization processing, and the normalized data is divided into data using a sliding window method to obtain a feature image.

[0022] Preferably, the calculation formula for the normalization process is:

[0023]

[0024] Where x is the original value, x max is a maximum value, x min is a minimum value.

[0025] Preferably, the wind speed prediction model includes a CNN module, an LSTM module and an Attention module, wherein the CNN module is used to extract time series features in the feature image;

[0026] The LSTM module captures the dependency between the time series features and generates information features;

[0027] The Attention module is used to divide the weights of information features to obtain wind speed prediction results.

[0028] Preferably, the CNN module includes a convolutional layer, a pooling layer and a fully connected layer, wherein the convolutional layer and the pooling layer are alternately stacked to deeply extract hidden features in the feature image, and the fully connected layer is used to reduce the complexity of the network.

[0029] Preferably, the CNN module adopts two-dimensional convolution.

[0030] Preferably, the LSTM module includes an input gate, a forget gate and an output gate, wherein the input gate is used to select important time series features, the forget gate is used to select unimportant time series features, and the output gate is used to output important time series features.

[0031] Preferably, the calculation process formula of the LSTM module is:

[0032] f t =σ(W f ·[h t-1 , x t ]+b f ).

[0033] i t =σ(W i ·[h t-1 , x t ]+b i ).

[0034]

[0035] Ot =σ(W o ·[h t-1 , x t ]+b o ).

[0036] h t =O t *tanh(C t );

[0037] Where x t represents the input vector; σ represents the sigmoid function; f t represents the output of the forget gate; W represents the weight value of different gate structures; b represents the bias term of different gate structures; h t-1 represents the hidden layer state; i t represents the input gate; C represents the state of the Cell; O t represents the output gate; h t Represents the output of the network.

[0038] Preferably, the calculation process formula of the Attention module is:

[0039]

[0040] α i =softmax(s(x i ,q));

[0041]

[0042] For n input information X=[x1,x2,…,x n ], q is the query vector, α i It indicates the relevance of the i-th information in the input information vector X to the query vector q when a query q is given. s represents the score function of the attention mechanism, and attention is the output of the attention mechanism.

[0043] Preferably, the mean absolute error and root mean square error of the wind speed prediction results are calculated to evaluate whether the results of the wind speed calculation model are accurate.

[0044] Due to the adoption of the above-mentioned technical solution, the present invention has the following beneficial effects:

[0045] The present invention discloses a method for predicting wind speed based on numerical weather forecast.

[0046] 1) Combining NWP data with actual wind turbine operation data can more accurately predict future wind speed conditions.

[0047] 2) The processing of outliers and missing values ​​ensures the quality of the data and avoids the negative impact of these data on model training; while data normalization ensures that data of different scales can be compared and analyzed within the same range, which helps the model learning process.

[0048] 3) The CNN module can effectively extract the spatial distribution features in the feature image, while the LSTM module is good at capturing temporal dependencies. The combination of the two can better understand the pattern of wind speed changes.

[0049] 4) The Attention mechanism can help the model focus on more important features, reduce the impact of irrelevant information, and increase the interpretability of the model.

[0050] 5) Calculating the mean absolute error (MAE) and root mean square error (RMSE) can quantitatively evaluate the accuracy of model predictions, which is helpful for model optimization and improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of the present invention;

[0052] Figure 2 It is a structural diagram of the CNN module;

[0053] Figure 3 It is a structural diagram of the LSTM module;

[0054] Figure 4 This is a workflow diagram of the CNN-LSTM-Attention prediction model;

[0055] Figure 5 Schematic diagram of the positional relationship between the target wind turbine and the four NWP nodes;

[0056] Figure 6 This is a data diagram of the normalized wind speed of the A2-2 wind turbine;

[0057] Figure 7 This is a trend diagram of the wind speed data time of the first 500 hours at node 1;

[0058] Figure 8 This is the thermal diagram of the correlation matrix between the wind speed of A2-2 fan and different wind speed data;

[0059] Fig. 9 A schematic diagram showing the comparison of the results of using different prediction models for the A2-2 unit;

[0060] Fig.10 A schematic diagram showing the comparison of the results of using different prediction models for the A3-5 unit;

[0061] Fig.11A schematic diagram showing the comparison of the results of using different prediction models for the A5-2 unit;

[0062] Fig.12 Schematic diagram of the comparison of results using different prediction models for the A5-3 unit. DETAILED DESCRIPTION

[0063] The technical solution of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. In the description, it should be understood that if there are terms such as "upper", "lower", "front", "back", "left", "right" and the like indicating directions or positional relationships, they are only corresponding to the drawings of the present invention, for the convenience of describing the present invention, and do not indicate or imply that the devices or components referred to must have a specific direction:

[0064] Embodiment 1 is:

[0065] Combined with Figure 1 The wind speed prediction method based on numerical weather forecast comprises the following specific steps:

[0066] (1) Collect NWP data of several grid nodes in the target wind farm and the operation data of wind turbines in the wind farm during the same period, and establish feature vectors;

[0067] In this embodiment, the wind farm selected for collection is located in North China, with the geographical coordinates of 114°21′ east longitude and 40°59′ north latitude. The local prevailing wind direction is northwest wind with an average wind speed of 7.27 m / s. The NWP data used is from the ERA5 reanalysis data set of the European Centre for Medium-Range Weather Forecasts, where the spatial resolution of the NWP data is 0.25° and the temporal resolution is 1 hour.

[0068] (2) Construct a wind speed prediction model;

[0069] The wind speed prediction model includes a CNN module, an LSTM module and an Attention module. Since CNN can extract spatial features in a sequence through a series of convolution and pooling operations, and the LSTM unit can model the feature sequence and capture the long-term dependencies in the sequence, by combining CNN and LSTM together, the CNN-LSTM network can simultaneously utilize spatial and temporal information to better process the structure and temporal relationship in the sequence data. Finally, different weights are given to data of different importance through the Attention mechanism to obtain the final wind speed prediction value.

[0070] The CNN module is used to extract the temporal features in the feature image;

[0071] The CNN model is a classic and widely used network structure in deep learning. Figure 2 As shown in the figure, it includes convolutional layers, pooling layers and fully connected layers. By alternately stacking convolutional layers and pooling layers, hidden features in the data can be deeply extracted. At the same time, due to its features such as local connection and weight sharing, it can effectively reduce the complexity of the network and is easy to train and optimize.

[0072] According to the length and feature position of the constructed time series, this CNN model uses two-dimensional convolution (2D-CNN) and two-dimensional pooling layer (MaxPooling2D), which can better extract local features in the wind speed data matrix, explore the mapping relationship between the wind speed data of each wind turbine and the wind speed data of different pressure layers of NWP, and finally send the extracted features to the LSTM module for processing.

[0073] The LSTM module captures the dependency between the time series features and generates information features;

[0074] Among them, the long short-term memory neural network (LSTM) is an improved recurrent neural network. Although it is designed to solve the problem that ordinary RNN recurrent neural networks cannot handle long-distance dependencies, LSTM has a relatively excellent performance in dealing with timing problems and can effectively solve the gradient explosion or gradient disappearance problems of RNN.

[0075] LSTM consists of three parts: input gate, forget gate and output gate. Its structure is as follows Figure 3 As shown in the figure, the input gate is used to select important time series features, the forget gate is used to select unimportant time series features, and the output gate is used to output important time series features.

[0076] The relevant calculation formula is:

[0077] f t =σ(W f ·[h t-1 , x t ]+b f ).

[0078] i t =σ(W i ·[h t-1 , x t ]+b i ).

[0079]

[0080] O t =σ(W o ·[h t-1 , x t ]+b o ).

[0081] h t =O t *tanh(C t );

[0082] Where x t represents the input vector; σ represents the sigmoid function; f t represents the output of the forget gate; W represents the weight value of different gate structures; b represents the bias term of different gate structures; h t-1 represents the hidden layer state; i t represents the input gate; C represents the state of the Cell; O t represents the output gate; h t Represents the output of the network.

[0083] The Attention module is used to divide the weights of information features to obtain wind speed prediction results;

[0084] The Attention module is inspired by the fact that the human brain focuses on a specific situation, obtains information about the specific situation, and ignores other unimportant information. Therefore, adding an attention mechanism after the LSTM network can weight the information features output by the LSTM layer in the prediction model, extract key information, and improve the model accuracy.

[0085] The calculation formula of the Attention module is:

[0086]

[0087] α i =softmax(s(x i ,q));

[0088]

[0089] For n input information X=[x1,x2,…,x n ], q is the query vector, α i It indicates the relevance of the i-th information in the input information vector X to the query vector q when a query q is given. s represents the score function of the attention mechanism, and attention is the output of the attention mechanism.

[0090] (3) The feature vector is converted into a feature image after preprocessing;

[0091] The preprocessing process includes outlier processing, missing value processing and data normalization processing, and the normalized data is divided into data using a sliding window method to obtain a feature image;

[0092] The calculation formula for normalization is:

[0093]

[0094] Where x is the original value, x max is a maximum value, x min is a minimum value.

[0095] (4) Input the feature image into the wind speed prediction model to obtain the wind speed prediction result;

[0096] In addition, by calculating the mean absolute error and root mean square error of the wind speed prediction results, it can be used to evaluate whether the results of the wind speed calculation model are accurate.

[0097] In addition, combined with Figure 4 In a specific embodiment of the present invention, wind speed data of four NWP grid nodes near the target wind farm are selected, and the node coordinates are: node 1 (41°N, 114.25°E), node 2 (41°N, 114.5°E), node 3 (40.75°N, 114.25°E) and node 4 (40.75°N, 114.5°E); the selected node data include the wind speed at a height of 100m at the four node positions and the wind speed data of seven pressure layers between 500hPa and 775hPa;

[0098] Considering the influence of terrain and geographical location on wind speed, in order to prove the effectiveness of wind speed prediction, four wind turbines (A2-2, A3-5, A5-2 and A6-3) in different areas and terrains in the wind farm were selected to carry out the experiment. The position relationship between the four wind turbines in the target wind farm and the four NWP nodes is shown in the figure below: Figure 5 As shown;

[0099] The NWP data of four grid nodes near the target wind farm were downloaded from the ERA5 dataset. The selected time period was from 0:00 on January 1, 2021 to 23:00 on July 31, 2021, with a total of 5088 hours of data.

[0100] In addition, data collection was carried out on the SCADA (Supervisory Control and Data Acquisition) system of the four wind turbines A2-2, A3-5, A5-2 and A6-3 in the wind farm. The wind speed of the wind turbines at one-minute intervals between January 1, 2021 and July 31, 2021 was selected and converted into data with a time resolution of one hour. The SCADA data and the NWP data were spliced ​​into a set of time series. Data processing was performed on them, abnormal data was screened and removed, and missing values ​​were interpolated or replaced. In order to facilitate the training of the prediction model, the data after removing the abnormal data was normalized.

[0101]

[0102] Where x is the original value, xmax is a maximum value, x min The normalized wind speed data of A2-2 fan is as follows Figure 6 As shown, Figure 6 The horizontal axis is the wind speed, and the vertical axis is the normalized result value.

[0103] The first 80% of the data in the data set are selected as the training set, and the last 20% of the data are selected as the test set. The data is divided using a sliding window method, that is, the first n points [x1, x2, ..., x n ] corresponds to point n+1, and so on to complete the data set division.

[0104] The input data are the SCADA operation data of the target wind turbine and the wind speed data of the four NWP grid nodes near the target wind farm. They are concatenated into vectors with a unified time resolution, and the sliding window length is set to 24. They are combined into a 24×33 feature vector matrix and input into the prediction model. The input structure of the prediction model is shown in Table 1.

[0105] Table 1 Prediction model data input structure

[0106] Table1 Structure of data inputs to the prediction model

[0107]

[0108] The trend of wind speed data over time in the first 500 hours at NWP node 1 is as follows: Figure 7 As shown, from Figure 7 It can be seen that the wind speed data at node 1 fluctuate greatly and the periodic characteristics are not obvious, but the time trends between the wind speed data are highly consistent. Therefore, it is feasible to predict the target wind turbine wind speed based on the transfer characteristics between NWP data.

[0109] In order to further reveal the correlation between different wind speed data at each node, the concept of Pearson correlation coefficient is introduced. Pearson correlation coefficient can measure the linear correlation between data and is an analytical indicator of the degree of correlation between data. Its formula is as follows:

[0110]

[0111] Where ρ (x，y) Represents the value of the Pearson correlation coefficient, which ranges from ±1. i and i represent the observed values ​​of node wind speed data and wind turbine data respectively, and and They represent their means, ρ (x，y)A value greater than 0 indicates a positive correlation, and a value less than 0 indicates a negative correlation. The closer the absolute value is to 1, the stronger the linear correlation between the data. Taking the A2-2 unit as an example, the A2-2 wind speed data is analyzed with the different wind speed data at the four nodes, and the correlation matrix heat map is drawn as follows: Figure 8 shown, and Figure 8 The horizontal axis is the wind speed data of each node at the location of the A2-2 unit, and the vertical axis is the predicted wind speed data of each node at the location of the A2-2 unit.

[0112] Depend on Figure 8 It can be seen that the wind speed of A2-2 wind turbine has a high correlation with the wind speed data of nodes 1 and 2, and is higher than the correlation between the wind speed of A2-2 wind turbine and nodes 3 and 4. The main reason is the limitation of NWP resolution, which is specifically manifested in that the target wind farm is closer to nodes 1 and 2. In addition, the correlation between wind speed data of different pressure layers at the same node is high, indicating that wind speed data of different pressure layers have transfer characteristics. For example, at node 1, the correlation coefficient between its 500hpa wind speed data and 550hpa wind speed data is 0.83, and the correlation coefficient between 550hpa wind speed data and 600hpa wind speed data is 0.70. The correlation coefficients of the wind speed data of the remaining pressure layers and the wind speed data of the higher-level pressure layers are all greater than 0.69.

[0113] Finally, the experiment uses two indicators, mean absolute error (MAE) and root mean square error (RMSE), to evaluate the effect of the model. The formula is as follows:

[0114]

[0115] where y t is the node wind speed data at time t, is the model predicted wind speed data at time t, and n is the total number of samples.

[0116] In addition, comparative experiments were carried out by inputting the fused feature data of four target wind turbines, A2-2, A3-5, A5-2 and A6-3, into the LSTM, 2DCNN, CNN-LSTM and CNN-LSTM-Attention models respectively, and the sliding window method was used to predict the wind speed data of the target wind turbines for the next 72 hours.

[0117] The prediction results of 4 units under different prediction models are as follows: Figure 9-12 Table 2 shows the calculation results of the evaluation indicators corresponding to the prediction experiment:

[0118] Table 2 Comparison of wind speed prediction performance evaluation indicators of the four units

[0119] Table 2 Comparison of evaluation indexes of wind speed prediction performance of 4 turbines

[0120]

[0121] And from Figure 9-12 It can be seen that the prediction effect of a single LSTM model is average. Its gating mechanism can capture the overall trend of the wind speed sequence, but it performs poorly for local wind speed sequences with strong fluctuations and cannot accurately predict. The 2DCNN model can extract latent features, making the overall trend of the curve closer to the actual wind speed sequence, but there is still a large gap between the wind speed prediction at a specific time point and the actual value. In contrast, the CNN-LSTM model inherits the advantages of both, and both the overall trend and the prediction at a specific time point are greatly improved compared to a single model. This model uses the Attention mechanism based on the CNN-LSTM model to improve the model's focus on important features, making it more able to fit the actual wind speed curve and improve the prediction accuracy.

[0122] Table 2 shows the performance comparison results of the four target wind turbines under different prediction models. It can be seen that the CNN-LSTM-Attention model has the smallest MAE and RMSE errors.

[0123] The CNN-LSTM-Attention model has the highest prediction accuracy in terms of performance. In the four groups of experiments, the MAE of the CNN-LSTM-Attention model was reduced by up to 52.82% compared with the LSTM model, by up to 41.04% compared with the 2DCNN model, and by up to 24.44% compared with the CNN-LSTM model; the RMSE of the CNN-LSTM-Attention model was reduced by up to 47.95% compared with the LSTM model, by up to 28.51% compared with the 2DCNN model, and by up to 23.06% compared with the CNN-LSTM model;

[0124] In summary, compared with other models, the CNN-LSTM-Attention prediction model proposed in this model has the best wind speed prediction effect for the four target units.

[0125] The parts of the present invention that are not described in detail are prior art. It is obvious to those skilled in the art that the present invention is not limited to the details of the above-mentioned exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention; therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is limited by the appended claims rather than the above description. Therefore, it is intended to include all changes within the meaning and scope of the equivalent elements of the claims in the present invention, and any figure marks in the claims should not be regarded as limiting the claims involved.

Claims

1. A wind speed prediction method based on numerical weather forecast, characterized in that: The specific steps are: Collect NWP data of several grid nodes in the target wind farm, as well as the operation data of wind turbines in the wind farm during the same period, and establish feature vectors; Construct wind speed prediction model; The feature vector is converted into a feature image after preprocessing; Input the feature image into the wind speed prediction model to obtain the wind speed prediction result.

2. The method for predicting wind speed based on numerical weather forecast according to claim 1, characterized in that: The preprocessing process includes outlier processing, missing value processing and data normalization processing, and the normalized data is divided into data using a sliding window method to obtain a feature image.

3. The method for predicting wind speed based on numerical weather forecast as claimed in claim 2, characterized in that: The calculation formula for the normalization process is: Where x is the original value, x max is a maximum value, x min is a minimum value.

4. The method for predicting wind speed based on numerical weather forecast according to claim 1, characterized in that: The wind speed prediction model includes a CNN module, an LSTM module and an Attention module, wherein the CNN module is used to extract time series features in the feature image; The LSTM module captures the dependency between the time series features and generates information features; The Attention module is used to divide the weights of information features to obtain wind speed prediction results.

5. The method for predicting wind speed based on numerical weather forecast as claimed in claim 4, characterized in that: The CNN module includes a convolutional layer, a pooling layer and a fully connected layer, wherein the convolutional layer and the pooling layer are alternately stacked to deeply extract hidden features in the feature image, and the fully connected layer is used to reduce the complexity of the network.

6. The method for predicting wind speed based on numerical weather forecast as claimed in claim 5, characterized in that: The CNN module adopts two-dimensional convolution.

7. The method for predicting wind speed based on numerical weather forecast as claimed in claim 4, characterized in that: The LSTM module includes an input gate, a forget gate and an output gate, wherein the input gate is used to select important time series features, the forget gate is used to select unimportant time series features, and the output gate is used to output important time series features.

8. The method for predicting wind speed based on numerical weather forecast as claimed in claim 4, characterized in that: The calculation process formula of the LSTM module is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) The t =σ(W o ·[h t-1 ,x t ]+b o ) h t =O t *tanh(C t ); Where x t represents the input vector; σ represents the sigmoid function; f t represents the output of the forget gate; W represents the weight value of different gate structures; b represents the bias term of different gate structures; h t-1 represents the hidden layer state; i t represents the input gate; C represents the state of the Cell; O t represents the output gate; h t Represents the output of the network.

9. The method for predicting wind speed based on numerical weather forecast as claimed in claim 4, characterized in that: The calculation formula of the Attention module is: α i =softmax(s(x i ,q)); For n input information X=[x1,x2,…,x n ], q is the query vector, α i It indicates the relevance of the i-th information in the input information vector X to the query vector q when a query q is given. s represents the score function of the attention mechanism, and attention is the output of the attention mechanism.

10. The method for predicting wind speed based on numerical weather forecast according to claim 1, characterized in that: The mean absolute error and root mean square error of the wind speed prediction results are calculated to evaluate whether the results of the wind speed calculation model are accurate.

Images