Coastal wind speed forecasting method and system based on deep learning

Abstract

The present invention relates to a coastal wind speed forecasting method and system based on deep learning, which belongs to the field of weather forecasting and includes the following steps: (1) establishing a wind speed sample set and a forecasting factor; (2) establishing and screening a BLSTM-TRA wind speed forecasting model based on the data of step (1); (3) training the BLSTM-TRA wind speed forecasting model; (4) testing the BLSTM-TRA wind speed forecasting model; (5) constructing a total loss function of the BLSTM-TRA wind speed forecasting model. The advantages of the present invention are: based on the advantages of the long short-term memory network (LSTM) method in solving nonlinear complex relationships, a BiLSTM model is used to process wind speed data, and the bidirectional cyclicity of BiLSTM is used to process the temporal nature of the data. The ability of Transformer to learn global information is utilized, and the multi-head attention mechanism of Transformer is used to extract effective features from wind speed data and train the model. By constructing the BLSTM-TRA model, a new algorithm and technology is provided for wind speed forecasting in coastal areas to further improve the level of wind speed forecasting in coastal areas.

Description

Technical Field

[0001] The invention relates to a coastal wind speed forecasting method based on deep learning, and belongs to the field of meteorological forecasting. Background Art

[0002] Establishing a weather forecast system not only involves the collection and storage of climate, geography, environment and other data, but most importantly, making accurate forecasts based on the acquired data. Strong winds are one of the most important weather products and play a key role in many forecasting applications. Strong wind forecasts are affected by multiple factors such as climate background and weather type. They are one of the atmospheric quantities that are difficult to accurately predict using numerical models. Improving the accuracy of strong wind forecasts is crucial to improving the forecast results of the entire weather forecast system.

[0003] In recent years, with the development of observation systems and forecasting technology, gale forecasting has made great progress, but how to further improve the level of wind speed prediction is still a problem. A series of studies have shown that the accuracy of gale forecasting is not simply related to longitude and latitude, but involves various factors such as terrain. Therefore, it is necessary to further explore the relationship between influencing factors such as temperature and humidity and gale, and fully consider any factors that may affect the results. In existing artificial intelligence technology, wind speed prediction is mainly based on algorithms based on single models. These methods can effectively predict wind speed in the short term, but the long-term dependence problem is less effective. The current deep learning method can achieve long-term prediction of wind speed, but the accuracy needs to be improved. Summary of the invention

[0004] In order to overcome the defects of the prior art, the present invention provides a coastal wind speed forecasting method based on deep learning. The technical solution of the present invention is:

[0005] A coastal wind speed forecasting method based on deep learning comprises the following steps:

[0006] (1) Establish wind speed sample set and prediction factors;

[0007] (2) Establish and screen the BLSTM-TRA wind speed forecast model based on the data in step (1);

[0008] (3) Training the BLSTM-TRA wind speed forecast model;

[0009] (4) Test the BLSTM-TRA wind speed forecast model;

[0010] (5) Construct the total loss function of the BLSTM-TRA wind speed forecast model.

[0011] The step (1) is specifically as follows:

[0012] 1.1: Select the coastal site that needs to be forecasted, obtain the historical data of local wind speed, use wind speed as the label of model input, and obtain meteorological data including precipitation, temperature, relative humidity, air pressure and wind direction;

[0013] 1.2: Establish a prediction factor data set. The prediction factors for each station include: precipitation, temperature, relative humidity, air pressure, wind direction, and wind speed;

[0014] 1.3: Clean the abnormal wind speed data of the sample set. The abnormal wind speed data is determined by calculating the difference between the wind speed data and the data at adjacent time points and determining the difference threshold. If the absolute value of the difference is greater than the threshold, the data is considered abnormal;

[0015] 1.4: Perform normalization processing and convert it into a dimensionless relative quantity, normalize its numerical range to the range of [0,1], and use the normalized prediction factors and labels as the prediction factors and wind speed sample set.

[0016] The step (2) is specifically as follows:

[0017] 2.1: The forecast factors and wind speed data sample sets in step 1.2 are divided into training set, validation set and test set in a ratio of 10:1:1. The model is trained by using the sliding window training method. According to the different wind speed forecast model characteristics of each station, different BLSTM-TRA wind speed forecast models are trained. The input end is various forecast factors, and the output end is wind speed. The nonlinear relationship between the forecast factors and wind speed is established;

[0018] 2.2: The model optimizer selects the Adam optimization algorithm, and uses the estimation of the first-order moment and the second-order moment to adjust the learning rate to adapt to the changes in different parameters and gradients; the loss function uses the mean square loss function for gradient descent iterative training:

[0019] Among them, y i is the true value of the i-th sample, is the predicted value of the i-th sample, n is the number of samples; the model training speed is improved by using the mini-batch gradient descent method.

[0020] The step (3) is specifically as follows:

[0021] 3.1. Input data of wind speed field at each moment x = (x1, x2, ..., x t ), first processed by BLSTM: x LSTM ,h LSTM =LSTM(x); get the feature vector x of the wind speed field in the high-dimensional feature space at each moment LSTM , LSTM(x) means LSTM calculation, xLSTM is the input data of Transformer, h LSTM is the hidden layer output of LSTM; the BLSTM layer provides the LSTM hidden state representation, and then x LSTM Send to Transformer module;

[0022] 3.2. For x LSTM To encode each feature in x LSTM The position information of each feature in the sequence is encoded to save the relative or absolute position of the wind speed at each moment in the sequence, ensuring that the sequential structure is not lost after the data is input into the model. The formula for position embedding is:

[0023]

[0024] In the formula, k represents the dimension, d m is the input dimension of the model, and pos represents the position of the input data in the sequence;

[0025] 3.3. Encode the data using an encoder;

[0026] 3.4. The output of the encoder is used as the input of the decoder. In the decoder, the input data is first subjected to multi-head self-attention operation, and the degradation problem of the multi-layer network is solved through residual connection. Then the input of each layer of neurons is normalized, and finally the output dimension of the data is adjusted through the fully connected layer to obtain the prediction result y = (y t+1 ,y t+2 , ..., y t+6 ).

[0027] The step (4) is specifically as follows:

[0028] The input data of the model is the complete meteorological element field of the past 24 time steps. Through the BLSTM module, the input wind speed field is encoded into the feature vector of the wind speed field in the high-dimensional feature space at each moment; then, the feature vectors of the high-dimensional wind speed field at the past 24 moments are input into the feature extraction module of the Transformer. The Transformer module establishes the long-term dependency relationship between the feature vectors of the wind speed field, and uses this to deduce the feature vectors of the wind speed field prediction results for the next 6 moments; finally, it is decoded into a complete wind speed field and the prediction results are output.

[0029] The step (5) is specifically as follows: using the predicted wind speed sequence y pred and the true wind speed series y true To determine the accuracy of the prediction, the formula is

[0030] A system for implementing a deep learning coastal wind speed forecasting method, comprising:

[0031] The data collection and preprocessing module is used to collect wind speed sample data in coastal areas and preprocess the collected data;

[0032] Feature selection and construction module, used to select prediction factors related to wind speed forecast from the preprocessed data;

[0033] Model building and screening module, used to build wind speed forecast model based on bidirectional long short-term memory network BLSTM and TRA architecture;

[0034] Model training module, used to train the BLSTM-TRA model;

[0035] Model testing module, used to test the trained model;

[0036] The loss function builds a module that evaluates the difference between the model's predictions and the actual wind speed values.

[0037] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0038] A computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0039] The advantages of the present invention are:

[0040] Based on the advantages of the long short-term memory network (LSTM) method in solving nonlinear complex relationships, the BiLSTM model is used to process wind speed data, and the bidirectional loop of BiLSTM is used to process the temporal nature of the data. Taking advantage of the Transformer's ability to learn global information, the Transformer's multi-head attention mechanism is used to extract effective features from the wind speed data and train the model. Through the actual observation data of the meteorological observation station in the coastal area, the nonlinear relationship between the wind speed in the coastal area and other related forecasting factors is established, and the BLSTM-TRA model is constructed, which provides a new algorithm and technology for the wind speed forecast in the coastal area, so as to further improve the level of wind speed forecast in the coastal area. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the main structural block diagram of the system of the present invention.

[0042] Figure 2 It is a test set experimental result diagram of the present invention.

[0043] Figure 3 It is a flow chart of wind speed prediction of the present invention.

[0044] Figure 4 It is the wind speed value of the gantry crane 67 in February predicted by the neural network in the embodiment of the present invention.

[0045] Figure 5 It is the wind speed value of the gantry crane 46 in February predicted by the neural network of the embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with specific embodiments, and the advantages and features of the present invention will become clearer as the description proceeds. However, these embodiments are exemplary only and do not constitute any limitation to the scope of the present invention. It should be understood by those skilled in the art that the details and forms of the technical solution of the present invention may be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the scope of protection of the present invention.

[0047] See also Figures 1 to 5 , the present invention relates to a coastal wind speed forecasting method based on deep learning, comprising the following steps:

[0048] (1) Establish wind speed sample set and prediction factors;

[0049] (2) Establish and screen the BLSTM-TRA wind speed forecast model based on the data in step (1);

[0050] (3) Training the BLSTM-TRA wind speed forecast model;

[0051] (4) Test the BLSTM-TRA wind speed forecast model;

[0052] (5) Construct the total loss function of the BLSTM-TRA wind speed forecast model.

[0053] Based on the above steps, we have achieved:

[0054] High-precision forecasting: Using deep learning models, especially the one combining Bidirectional Long Short-Term Memory (BLSTM) and Transformer (TRA), can capture the complex patterns and long-term dependencies of wind speed changes, thus improving the accuracy of forecasts.

[0055] Automated processing: Automatically establish sample sets and predictive factors, reduce manual intervention, and improve the efficiency and accuracy of data processing.

[0056] Model screening: Through the screening process, the optimal model structure and parameters can be selected to ensure the best performance of the model in practical applications.

[0057] Powerful data processing capabilities: The BLSTM-TRA model is able to process large amounts of historical wind speed data and other relevant forecasting factors, thereby improving the reliability of the forecast.

[0058] The step (1) is specifically as follows:

[0059] 1.1: Select the coastal site that needs to be forecasted, obtain the historical data of local wind speed, use wind speed as the label of model input, and obtain meteorological data including precipitation, temperature, relative humidity, air pressure and wind direction;

[0060] 1.2: Establish a prediction factor data set. The prediction factors for each station include: precipitation, temperature, relative humidity, air pressure, wind direction, and wind speed;

[0061] 1.3: Clean the abnormal wind speed data of the sample set. The abnormal wind speed data is determined by calculating the difference between the wind speed data and the data at adjacent time points, and determining the difference threshold. If the absolute value of the difference is greater than the threshold, the data is considered abnormal; common examples include null values ​​and garbled characters in wind speed data.

[0062] 1.4: Due to the large difference in dimensions between different variables, there will be a large difference. If they are input into the neural network model at the same time, the network will be overfitted. Therefore, normalization is performed to convert them into dimensionless relative quantities, and their numerical range is normalized to the range of [0,1]. The normalized forecast factors and labels are used as forecast factors and wind speed sample sets.

[0063] The step (2) is specifically as follows:

[0064] 2.1: The forecast factors and wind speed data sample sets in step 1.2 are divided into training set, validation set and test set in a ratio of 10:1:1. The model is trained by using the sliding window training method. According to the different wind speed forecast model characteristics of each station, different BLSTM-TRA wind speed forecast models are trained. The input end is various forecast factors, and the output end is wind speed. The nonlinear relationship between the forecast factors and wind speed is established;

[0065] 2.2: The model optimizer selects the Adam optimization algorithm, and uses the estimation of the first-order moment and the second-order moment to adjust the learning rate to adapt to the changes in different parameters and gradients; the loss function uses the mean square loss function for gradient descent iterative training:

[0066] Among them, y i is the true value of the i-th sample, is the predicted value of the i-th sample, n is the number of samples; the model training speed is improved by using the mini-batch gradient descent method.

[0067] 4. The coastal wind speed forecasting method based on deep learning according to claim 3, characterized in that the step (3) is specifically:

[0068] 3.1. Input data of wind speed field at each moment x = (x1, x2, ..., x t ), first processed by BLSTM: x LSTM ,h LSTM =LSTM(x); get the feature vector x of the wind speed field in the high-dimensional feature space at each moment LSTM , LSTM(x) indicates LSTM calculation, xLSTM is the input data of Transformer, and hLSTM is the hidden layer output of LSTM; the BLSTM layer provides the LSTM hidden state representation, and then sends xLSTM to the Transformer module;

[0069] 3.2. Since the order relationship between data is very important for time series prediction problems, and Transformer uses global information instead of the time series structure in the recurrent neural network, it is required that in addition to encoding each feature in xLSTM, the position information of each feature in xLSTM must also be encoded (i.e., position embedding), and the relative or absolute position of the wind speed at each moment in the sequence must be saved to ensure that the order structure is not lost after the data is input into the model. The formula for position embedding is:

[0070]

[0071] In the formula, k represents the dimension, dm is the input dimension of the model, and pos represents the position of the input data (wind speed, time, temperature, humidity, etc.) in the sequence;

[0072] 3.3. Use the encoder to encode the data; the encoder consists of two submodules: the multi-head self-attention module and the feedforward neural network module. The multi-head self-attention module includes multiple self-attention layers, which can simultaneously focus on different positions in the input sequence, thereby capturing the relationship between sequences, and then passing the data to the feedforward neural network module, and the encoder outputs an intermediate result.

[0073] 3.4. The output of the encoder is used as the input of the decoder. In the decoder, the input data is first subjected to multi-head self-attention operation, and the degradation problem of the multi-layer network is solved through residual connection. Then the input of each layer of neurons is normalized, and finally the output dimension of the data is adjusted through the fully connected layer to obtain the prediction result y = (y t+1 ,y t+2 , ..., y t+6 ).

[0074] The step (4) is specifically as follows:

[0075] The input data of the model is the complete meteorological element field of the past 24 time steps. Through the BLSTM module, the input wind speed field is encoded into the feature vector of the wind speed field in the high-dimensional feature space at each moment; then, the feature vectors of the high-dimensional wind speed field at the past 24 moments are input into the feature extraction module of the Transformer. The Transformer module establishes the long-term dependency relationship between the feature vectors of the wind speed field, and uses this to deduce the feature vectors of the wind speed field prediction results for the next 6 moments; finally, it is decoded into a complete wind speed field and the prediction results are output.

[0076] The step (5) is specifically as follows: using the predicted wind speed sequence y pred and the true wind speed series y true To determine the accuracy of the prediction, the formula is

[0077] A system for implementing a deep learning coastal wind speed forecasting method, comprising:

[0078] Data collection and preprocessing module 1, used to collect wind speed sample data in coastal areas and preprocess the collected data;

[0079] The data collection module is used to obtain wind speed data, including a historical wind speed data acquisition submodule and a meteorological data acquisition submodule. The historical wind speed data acquisition submodule is used to obtain historical data of local wind speed, and the meteorological data acquisition submodule is used to obtain meteorological data, including ground meteorological elements: precipitation, temperature, relative humidity, air pressure, wind direction, 10m radial wind and 10m latitudinal wind, 2m temperature, 2m dew point temperature, and average sea level pressure; high-altitude meteorological elements include: 500hPa, 700hPa, 850hPa, 925hPa relative humidity, 500hPa, 700hPa, 850hPa, 925hPa radial wind and latitudinal wind, 500hPa, 700hPa, 850hPa, 925hPa temperature.

[0080] The data preprocessing module is used to preprocess the data, including a data cleaning unit and a data integration unit. The data cleaning unit processes the null values ​​and garbled codes in the data matrix, and the data integration unit is used to integrate the data and standardize the data of each element to improve the operation efficiency.

[0081] Feature selection and construction module 2 is used to select prediction factors related to wind speed forecast from preprocessed data; this feature selection and construction module includes bidirectional long short-term memory (BiLSTM) neural network unit and Transformer unit. The complementarity of LSTM for long-term representation of wind speed time series data is utilized, and the characteristics of different weights are given to features through Transformer. Using BiLSTM to model wind speed time series data can handle long sequences and long-term dependency problems. Through bidirectional transmission, information from the past and the future can be used to improve wind speed prediction capabilities.

[0082] Model building and screening module 3, used to build a wind speed forecast model based on the bidirectional long short-term memory network BLSTM and TRA architecture;

[0083] Model training module 4, used to train the BLSTM-TRA model;

[0084] The model testing module 5 is used to test the trained model (to predict the wind speed); the loss function building module 6 is used to evaluate the difference between the model prediction value and the actual wind speed value.

[0085] The model test module includes a wind speed prediction module, a prediction result evaluation module and a result output module. The wind speed prediction module is specifically: the wind speed field x at each moment in the input data is (x1, x2, ..., x t ), first processed by BLSTM, the feature vector of the wind speed field in the high-dimensional feature space at each moment is obtained and sent to the Transformer module. Since the Transformer uses global information instead of the temporal structure in the recurrent neural network, this requires that in addition to encoding each feature in the feature vector, the position information of each feature in the feature vector must also be encoded (i.e., position embedding), saving the relative or absolute position of the wind speed at each moment in the sequence, and ensuring that the sequential structure is not lost after the data is input into the model. The data after position embedding is processed through the encoding submodule and the decoding submodule, and finally the output dimension of the data is adjusted by the fully connected layer to obtain the prediction result y=(y t+1 ,y t+2 , ..., y t+6 ).

[0086] Prediction result evaluation module: includes error calculation submodule and trend analysis submodule. The error calculation submodule is used to calculate the error between the predicted value and the actual observed value. The trend analysis submodule is used to analyze the trend between the predicted value and the actual observed value.

[0087] Result output module: includes prediction result output submodule and visualization submodule. The prediction result output submodule is used to output the model prediction results, and the visualization submodule is used to visualize the results.

[0088] The present invention relates to a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0089] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the method when executed by a processor.

[0090] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0091] In view of the advantages of the long short-term memory network (LSTM) method in solving nonlinear complex relationships, the present invention uses the BiLSTM model to process wind speed data, and uses the bidirectional circularity of BiLSTM to process the temporal nature of the data. Utilizing the ability of Transformer to learn global information, the multi-head attention mechanism of Transformer is used to extract effective features from wind speed data and train the model. Through the real-time observation data of the meteorological observation station in the coastal area, the nonlinear relationship between the wind speed in the coastal area and other related forecasting factors is established, and the BLSTM-TRA model is constructed, which provides a new algorithm and technology for the wind speed forecast in the coastal area, so as to further improve the level of wind speed forecast in the coastal area.

[0092] The following is a detailed description with examples:

[0093] The model was trained using the sliding window training method. According to the different wind speed forecast model characteristics of each station, the parameters were adjusted to achieve the best prediction effect. The model optimizer selected the Adam optimization algorithm, the learning rate was 0.01, and the loss function used the mean square loss function for gradient descent iterative training. The model training speed was improved by using the mini-batch gradient descent method.

[0094] In order to evaluate the performance of the algorithm, RMSE (root mean square error), MAE (mean absolute error) and CC (Correlation coefficient) are used as evaluation indicators. The model proposed by this algorithm is experimentally verified on the wind speed dataset in the southern coastal area of ​​Shandong Peninsula. The time range of the selected data is from 2006 to 2018, and the interval between adjacent moments is 1 hour. It is divided into three parts: training set (January 1, 2006 to December 31, 2016), validation set (January 1, 2017 to December 31, 2017) and test set (January 1, 2018 to December 31, 2018). The experimental results on the test set can be referred to Figure 2 . Using the method of the present invention, the predicted wind speed is better than that of the single LSTM method and the ECMWF method.

[0095] The minute data from September 2020 to February 2021 were selected, the validation set was the minute data from January 2021, and the test set was the minute data from February 2021. The blue is the actual wind speed, the red is the predicted wind speed, and the Figure 4 and Figure 5 It can be clearly seen that the wind speed can be better predicted by using the method of the present invention.

[0096] Table 1 shows the error comparison between the predictions of door machine 67 and door machine 46 in the test set.

[0097]

[0098]

[0099] Table 1

[0100] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A coastal wind speed forecasting method based on deep learning, characterized in that: The following steps are involved: (1) Establish wind speed sample set and prediction factors; (2) Establish and screen the BLSTM-TRA wind speed forecast model based on the data in step (1); (3) Training the BLSTM-TRA wind speed forecast model; (4) Test the BLSTM-TRA wind speed forecast model; (5) Construct the total loss function of the BLSTM-TRA wind speed forecast model.

2. The coastal wind speed forecasting method based on deep learning according to claim 1, characterized in that: The step (1) is specifically as follows: 1.1: Select the coastal site that needs to be forecasted, obtain the historical data of local wind speed, use wind speed as the label of model input, and obtain meteorological data including precipitation, temperature, relative humidity, air pressure and wind direction; 1.2: Establish a prediction factor data set. The prediction factors for each station include: precipitation, temperature, relative humidity, air pressure, wind direction, and wind speed; 1.3: Clean the abnormal wind speed data of the sample set. The abnormal wind speed data is determined by calculating the difference between the wind speed data and the data at adjacent time points and determining the difference threshold. If the absolute value of the difference is greater than the threshold, the data is considered abnormal; 1.4: Perform normalization processing and convert it into a dimensionless relative quantity, normalize its numerical range to the range of [0,1], and use the normalized prediction factors and labels as the prediction factors and wind speed sample set.

3. The coastal wind speed forecasting method based on deep learning according to claim 2, characterized in that: The step (2) is specifically as follows: 2.1: The prediction factors and wind speed data sample set in step 1.1 are divided into training set, validation set and test set in a ratio of 10:1:

1. The model is trained by using the sliding window training method. According to the different wind speed prediction model characteristics of each station, different BLSTM-TRA wind speed prediction models are trained. The input end is various prediction factors, and the output end is wind speed. The nonlinear relationship between prediction factors and wind speed is established; 2.2: The model optimizer selects the Adam optimization algorithm and uses the estimation of the first-order moment and the second-order moment to adjust the learning rate to adapt to the changes of different parameters and gradients; The loss function uses the mean square loss function for gradient descent iterative training: Among them, yi is the true value of the i-th sample, is the predicted value of the i-th sample, n is the number of samples; the model training speed is improved by using the mini-batch gradient descent method.

4. The coastal wind speed forecasting method based on deep learning according to claim 3 is characterized in that: The step (3) is specifically as follows: 3.

1. Input data of wind speed field at each moment x=(x1,x2,...,x t ), first processed by BLSTM: x LSTM ,h LSTM =LSTM(x); get the feature vector x of the wind speed field in the high-dimensional feature space at each moment LSTM , LSTM(x) means LSTM calculation, x LSTM is the input data of Transformer, h LSTM is the hidden layer output of LSTM; the BLSTM layer provides the LSTM hidden state representation, and then x LSTM Send to Transformer module; 3.

2. For x LSTM To encode each feature in x LSTM The position information of each feature in the sequence is encoded to save the relative or absolute position of the wind speed at each moment in the sequence, ensuring that the sequential structure is not lost after the data is input into the model. The formula for position embedding is: In the formula, k represents the dimension, d m is the input dimension of the model, and pos represents the position of the input data in the sequence; 3.

3. Encode the data using an encoder; 3.

4. The output of the encoder is used as the input of the decoder. In the decoder, the input data is first subjected to multi-head self-attention operation, and the degradation problem of the multi-layer network is solved through residual connection. Then the input of each layer of neurons is normalized, and finally the output dimension of the data is adjusted through the fully connected layer to obtain the prediction result y = (y t+1 ,y t+2 ,…y t+6 ).

5. The coastal wind speed forecasting method based on deep learning according to claim 4 is characterized in that: The step (4) is specifically as follows: The input data of the model is the complete meteorological element field of the past 24 time steps. Through the BLSTM module, the input wind speed field is encoded into the feature vector of the wind speed field in the high-dimensional feature space at each moment; then, the feature vectors of the high-dimensional wind speed field at the past 24 moments are input into the feature extraction module of the Transformer. The Transformer module establishes the long-term dependency relationship between the feature vectors of the wind speed field, and uses this to deduce the feature vectors of the wind speed field prediction results for the next 6 moments; finally, it is decoded into a complete wind speed field and the prediction results are output.

6. The coastal wind speed forecasting method based on deep learning according to claim 5, characterized in that: The step (5) is specifically as follows: using the predicted wind speed sequence y pred and the true wind speed series y true To determine the accuracy of the prediction, the formula is 7. A system for implementing the coastal wind speed forecasting method based on deep learning according to any one of claims 1 to 6, characterized in that: include: The data collection and preprocessing module is used to collect wind speed sample data in coastal areas and preprocess the collected data; Feature selection and construction module, used to select prediction factors related to wind speed forecast from the preprocessed data; Model building and screening module, used to build wind speed forecast model based on bidirectional long short-term memory network BLSTM and TRA architecture; Model training module, used to train the BLSTM-TRA model; Model testing module, used to test the trained model; The loss function builds a module that evaluates the difference between the model's predictions and the actual wind speed values.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images