Single-point-oriented wind speed medium and long term prediction model construction method

Abstract

The invention relates to a single-point-oriented wind speed medium and long term prediction model construction method, which comprises the following steps of: constructing a training data set through operations of extracting a numerical prediction result, filtering and sorting data and the like, and adding corresponding lattice point prediction information extracted from historical adjacent days and adjacent time into the training data set in the same way; the method comprises the steps that firstly, a high-quality training data set is constructed, then meteorological elements having significant influence on wind field forecasting are screened through a random forest algorithm to serve as input features, then an AU-Net convolutional neural network fused with an attention mechanism is constructed, and when a correction model is trained, only a to-be-corrected quantity serves as the input of the model, and a single-element correction model is trained; inputting meteorological elements based on the selected wind speed correction, training a multi-element correction model, and performing precision evaluation on the trained model to realize prediction of medium and long term wind speed and improvement of the accuracy of the wind speed regional medium and long term forecast correction model.

Description

Technical Field

[0001] The present invention provides a method for constructing a medium- and long-term wind speed prediction model for a single point, which relates to the fields of meteorological monitoring, data feature screening, deep learning algorithm development, and belongs to the field of model algorithm development. Background Art

[0002] High winds are a common, high-impact weather event that can trigger a variety of secondary disasters, such as damage to buildings, disruptions to transportation, and fires. High wind speeds typically exceed 17 meters per second, sometimes reaching over 30 meters per second. According to weather forecasting standards, a two-minute average wind speed of force 6 (greater than 10.8 m / s) is generally considered a high wind standard (excluding typhoons and thunderstorms). In meteorological observations, wind speeds at a station 10 meters above the surface are considered instantaneous winds, and high winds are defined as wind speeds of 17 m / s or greater, or visually estimated at force 8 or greater. These intense winds not only cause significant inconvenience to people's daily lives but also pose serious threats and challenges to transportation, energy supply, agricultural production, construction, the environment, and public health. Therefore, the importance of high wind forecasts and warnings is self-evident. They are crucial to the stable development of society and the economy, as well as the safety of life and property. High wind forecasts are not only an essential component of meteorological services but also a crucial basis for emergency management. Departments and relevant institutions can formulate emergency plans in advance based on strong wind forecast information, rationally allocate resources, and organize emergency rescue teams to ensure that they can respond to various emergencies quickly and effectively when strong winds occur.

[0003] The formation mechanism of strong winds is complex, involving multiple factors such as atmospheric dynamics, thermodynamics and topography. In recent years, through in-depth research on strong wind events such as severe convective weather and typhoons, scientists have revealed more key influencing factors and physical mechanisms. Traditional forecasting methods for strong winds mainly rely on statistical forecasting and numerical forecasting. Statistical forecasting is based on historical data and does not take into account the nonlinear characteristics of wind field changes. As the forecast time increases, its accuracy is greatly reduced. In recent decades, with the advancement of science and technology and the leapfrog development of computer technology, numerical forecasting technology has gradually developed into the main force of meteorological business forecasting. Numerical model forecasting relies on its powerful computing power to perform numerical simulations on the partial differential equations of fluid mechanics and thermodynamics, thereby predicting the atmospheric motion state and weather phenomena in a certain period of time in the future. However, when simulating atmospheric changes, numerical models are difficult to fully simulate the internal changes of the atmosphere due to the fluidity and variability of the atmosphere, the uncertainty of boundary conditions and initial conditions, and the defects of the parameterization scheme itself. Therefore, errors in the numerical forecast results are difficult to avoid, and the ability to deal with abnormal weather still needs to be enhanced. Therefore, statistical post-processing of numerical model forecasts is necessary, which can effectively improve the accuracy of the forecast.

[0004] Model error correction can be divided into two main categories: the first category is to start from the model itself, improve the quality of the initial field, optimize the boundary conditions, improve the forecast initial perturbation construction scheme, etc.; the other category is to post-process the forecast results based on statistical models to reduce the model's systematic deviation. Since the statistical post-processing correction method has low cost and high practical value, it has been widely used in meteorological business. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to improve the accuracy of wind speed forecast. The present invention provides a method for constructing a medium- and long-term wind speed prediction model for a single point. By extracting and processing meteorological elements, a high-quality multi-element training dataset is constructed. The random forest algorithm is used to screen meteorological elements that have a significant impact on wind field forecasts as input features. The AU-Net deep learning network architecture integrating the attention mechanism is constructed. The prediction effect of the revised model is trained and verified to achieve medium- and long-term prediction of single-point wind speed.

[0006] The technical solution of the present invention is specifically as follows: A method for constructing a medium- and long-term wind speed prediction model for a single point includes the following steps: Step 1: Extract the training set of numerical forecast results. After filtering and organizing the data, extract the gridded numerical forecast products according to the n*n grid. The extracted parameters include temperature, pressure, humidity, wind field and precipitation. All this information needs to be integrated into the neural network for correction. Then, the corresponding grid forecast information is extracted from the historical near-day and near-time using the same method. At the same time, the information of ground observation data is also added to the training set to construct a high-quality training data set. Step 2: Using domain knowledge and feature engineering, select meteorological factors that have a significant impact on wind farm forecasts as input features, and perform data processing to remove irrelevant and noise features. Step 3: During data preprocessing, the grid data is expanded and treated as data from the same site, and a wind field correction model is constructed using the random forest algorithm. Step 4: Train and test the wind speed and direction correction models for different forecast timeframes, evaluate the correction effects for different forecast timeframes, and analyze the results. Step 5: Calculate the importance of the input factors for the numerical forecast wind field correction, select more representative correction input factors, and analyze the wind speed correction effect to achieve an improvement in the correction effect compared to single-factor input; Step 6: Build an AU-Net convolutional neural network with an integrated attention mechanism, input the extracted training set into the neural network, and use the neural network's feedback system to continuously adjust the network structure to improve the model's accuracy in predicting other meteorological factors. To verify that multi-factor input improves the correction results, when training the correction model, first use only the input to be corrected as the model input and train a single-factor correction model. Then, based on the selected wind speed correction input meteorological factor, train a multi-factor correction model. Step 7: Under the premise of single factor input, without considering the forecast timeliness, a verification evaluation is conducted based on the AU-Net model and the basic U-Net wind speed correction model, the prediction accuracy of the model is evaluated, and the results after model comparison are analyzed to improve the prediction accuracy of the regional medium- and long-term wind speed forecast correction model.

[0007] Based on the random forest algorithm, it includes the following steps: 3.1. Assume that the size of the sample set is N, the number of decision trees in the algorithm is K, and the number of features of the sample is M; 3.2. Construction of training sample set: Based on the original sample set, K new sample sets are randomly extracted using the bagging idea. Each sample set corresponds to a decision tree to be trained. 3.3. Decision tree training: Utilize the idea of ​​feature subspace to perform the most reasonable attribute partitioning. That is, for M features, only m features are selected according to certain rules to participate in the most reasonable attribute partitioning, and a well-trained decision tree is obtained. 3.4. Construction of random forest model: For k training sets, obtain K decision tree models according to the method in step 3.3, and obtain the wind field correction model based on random forest through ensemble learning.

[0008] Step 4 is to use the random forest method to calculate the importance of the input factors for the numerical forecast wind field correction and select more representative correction input factors. The main steps of the test are as follows: (1) The training data is divided according to the forecast time, and the wind speed and wind direction training data sets with different forecast time are obtained.

[0009] (2) Build a wind field correction model based on random forest and train wind speed and wind direction correction models for different forecast time periods; (3) Test the effectiveness of the wind speed and direction correction model on the validation set, and adjust the training parameters, including the maximum depth of the tree and the number of trees, based on the validation set results; (4) A wind field correction test was conducted on the test set to evaluate the correction effects of different forecast timeliness and calculate the importance of input factors. The wind speed and wind direction correction models were both based on a regression model with mean square error as the loss. For the wind speed correction model, the number of trees was set to 50 and the maximum depth was set to 10. For the wind direction correction model, the number of trees was set to 60 and the maximum depth was set to 13.

[0010] Step five is as follows: for the calculation of feature importance, a purity index based on a decision tree is selected and normalized; for the importance calculation results based on random forest, input factors with importance results greater than or equal to 0.05 are used as input variables for subsequent multi-factor correction, and input factors with importance results less than or equal to 0.05 are eliminated as "noise".

[0011] Therefore, based on the existing research on high wind forecasting technology, the present invention proposes a method for constructing a medium- and long-term wind speed forecast model for a single point. This method constructs an enhanced training dataset containing historical adjacent spatiotemporal information, combines the random forest algorithm to realize dynamic feature importance evaluation, and designs an AU-Net convolutional neural network architecture that integrates the attention mechanism. The method aims to break through the bottlenecks of traditional methods in data utilization, feature extraction and model generalization capabilities, and provide a new technical solution for improving the accuracy of medium- and long-term wind speed forecasts for a single point. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is the flow chart for the medium- to long-term wind farm forecast processing; Figure 2 Revised flow chart for wind farm; Figure 3 This is the wind speed correction map based on random forest; Figure 4 This is a wind speed comparison chart; Figure 5 This is the wind direction correction map based on random forest; Figure 6 This is a wind direction comparison chart; Figure 7 Corrected feature importance calculations for wind speed; Figure 8 Correct the feature importance calculation results for wind direction; Figure 9 AU-Net model structure; Figure 10 It is the Attention Gate structure of the attention module; Figure 11 Correction error change for wind speed; Figure 12 is the seasonal variation of the error. Specific implementation plan

[0013] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the implementation described is only a part of the implementation of the present invention, not all embodiments. Based on the implementation of the present invention, all other implementations obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The present invention proposes a method for constructing a medium- and long-term wind speed prediction model for a single point. A training data set is constructed by extracting the results of numerical forecasts and filtering and collating the data. The grid forecast information obtained from historical near-days and near-times is extracted in the same way, and the extracted data is added to the training data set to construct a high-quality training data set. Then, a random forest algorithm is used to screen meteorological factors that have a significant impact on wind field forecasts as input features, and then an AU-Net convolutional neural network integrating an attention mechanism is constructed. At the same time, in order to verify the improvement of the model correction results by multi-factor input, when training the correction model, only the amount to be corrected is used as the input of the model to train a single-factor correction model; then, based on the selected wind speed correction input meteorological factors, including 10m wind speed, surface pressure and 2m dew point temperature, a multi-factor correction model is trained. Then, the accuracy of the trained model is evaluated and the results are analyzed, thereby achieving an improvement in the accuracy of the medium- and long-term wind speed prediction and the regional medium- and long-term wind speed forecast correction model.

[0015] To achieve the above objectives, the specific steps of the present invention are as follows: Step 1: Extract the training set of numerical forecast results. After filtering and organizing the data, extract the gridded numerical forecast products according to the n*n grid. The extracted parameters include temperature, pressure, humidity, wind field and precipitation information. All this information needs to be integrated into the neural network for correction; then use the same method to extract the corresponding grid forecast information from the historical near-day and near-time. At the same time, the information of the ground observation data is added to the training set to construct a high-quality training data set. The extraction method is the same as the numerical forecast results. It should be noted that the resolution and forecast results must be unified.

[0016] Step 2: Through domain knowledge and feature engineering, meteorological factors that significantly influence wind forecasts are selected as input features. This method selects 16 wind-related output variables as basic input factors for wind correction, including historical wind speed and direction data, as well as related data such as temperature, humidity, and air pressure. The specific factor information is shown in Table 1. The extracted variables Uwind and Vwind are vector-synthesized into wind speed and direction, which are then used as input for subsequent experiments. During data processing, irrelevant and noisy features are removed to reduce feature dimensionality and improve the generalization and stability of the prediction model.

[0017] Table 1. Input elements of wind field forecast model Serial number Feature name 1 10m U wind (10metre U wind component) 2 10mV wind (10metreVwindcomponent) 3 925hPa pressure layer U wind (U component of wind: level 925) 4 925hPa pressure layer V wind (Vcomponentofwind:level925) 5 850hPa pressure layer U wind (U component of wind: level 850) 6 850hPa pressure layer V wind (Vcomponentofwind:level850) 7 2m temperature (2metretemperature) 8 925hPa pressure layer temperature (Temperature: level925) 9 850hPa pressure layer temperature (Temperature: level 850) 10 Relative humidity at 925hPa pressure layer (Relativehumidity: level925) 11 Relative humidity at 850hPa pressure layer (Relativehumidity: level850) 12 Geopotential height of the 925hPa pressure layer (GeopotentialHeight: level925) 13 Geopotential Height of the 850hPa pressure layer (Geopotential Height: level850) 14 2m dew point temperature (2metredewpointtemperature) 15 Surface pressure 16 Total Cloud Cover Step 3: During data preprocessing, the grid data is expanded and treated as data from a single site. This means that only one model is built for each forecast timeframe across the entire region. The inputs are the 16 meteorological elements from the selected numerical forecast output, and the labels are the wind field labels at the corresponding timeframes.

[0018] The main construction process based on the random forest algorithm is as follows Figure 2 As shown, it mainly includes the following steps: 3.1. Assume that the size of the sample set is N, the number of decision trees in the algorithm is K, and the number of features of the sample is M; 3.2. Construction of training sample set: Based on the original sample set, K new sample sets are randomly extracted using the bagging idea. Each sample set corresponds to a decision tree to be trained. 3.3. Decision tree training: Utilize the idea of ​​feature subspace to perform the most reasonable attribute partitioning. That is, for M features, only m features are selected according to certain rules to participate in the most reasonable attribute partitioning, and a well-trained decision tree is obtained. 3.4. Construction of random forest model: For k training sets, obtain K decision tree models according to the method in step 3.3, and obtain the wind field correction model based on random forest through ensemble learning.

[0019] Step 4: Use the random forest method to build a wind field correction model, train and test the wind speed and wind direction correction models for different forecast time periods, evaluate the correction effects of different forecast time periods and analyze the results. Specifically: Use the random forest method to calculate the importance of the input factors of the numerical forecast wind field correction and select more representative correction input factors. The numerical forecast data for 2023 and the corresponding wind field label data are used as the training set (including the validation set), and the numerical forecast data for 2024 and the corresponding wind field label data are used as the test set. The main steps of the test are as follows: (1) The training data is divided according to the forecast time, and the wind speed and wind direction training data sets with different forecast time are obtained.

[0020] (2) According to the attached Figure 2 A wind field correction model based on random forest is built to train wind speed and wind direction correction models for different forecast time periods.

[0021] (3) Test the effectiveness of the wind speed and direction correction model on the validation set, and adjust the training parameters, including the maximum depth of the tree and the number of trees, based on the validation set results.

[0022] (4) Wind field correction tests were conducted on the test set to evaluate the effects of different forecast time corrections and calculate the importance of input factors. It is worth mentioning that in order to better correct the numerical forecast results, the wind speed and wind direction correction models both use regression models with mean square error as the loss. For the wind speed correction model, the number of trees is set to 50 and the maximum depth is set to 10. For the wind direction correction model, the number of trees is set to 60 and the maximum depth is set to 13.

[0023] Correction effect of wind speed Figure 3 As shown, the left figure is the root mean square error of wind speed before and after correction, and the right figure is the mean absolute error. The solid line is the distribution of the error before correction with the forecast time, and the dotted line is the distribution of the error after correction with the forecast time. It can be seen from the figure that the correction model based on random forest has an overall stable correction effect on wind speed. The wind speed forecast error of each time period has been reduced, and as the forecast time period increases, the overall correction amount shows an increasing trend. In order to more intuitively feel the wind speed regional distribution before and after correction, the wind speed forecast results in the selected area of ​​​​the RMAPS model three hours after the forecast at 6:00 on May 30, 2023, as well as the corresponding correction results and ERA5 (label) results are shown in the attached figure. Figure 4 It can be seen from the figure that although the correction results are better than the numerical forecast results, some regional errors have obviously increased. On the whole, the original characteristics of the numerical forecast are maintained. This is also because the direct regional correction based on the station correction idea will inevitably lead to global changes. This process obviously cannot learn the characteristics of the real wind speed distribution, which limits the effect of the correction.

[0024] The correction effect of wind direction is shown in the attached Figure 5 As shown in the attached Figure 5 It can be seen from the figure that the error reduction of wind direction in the region by using random forest is much smaller than that of wind speed. In some cases, the effect after time-sensitive correction is even worse than the numerical forecast result. The root mean square error is not ideal before 9 hours. Figure 6 They are the wind direction results in the selected area of ​​the RMAPS model three hours after the forecast at 6:00 on May 30, 2023, as well as the corresponding correction results and ERA5 (label) results. The wind direction correction effect is obviously very poor, especially near the north wind (around 0° and 360°).

[0025] Step 5: Calculate the importance of the input factors for the correction of the numerical forecast wind field, select more representative correction input factors, and analyze the wind speed correction effect to achieve an improved correction effect compared to single-factor input.

[0026] Specifically: For the calculation of feature importance, we selected the purity index based on decision tree and performed normalization. The results are shown in the attached figure. Figure 7 and attached Figure 8 As shown in the figure, for the importance calculation results based on random forests, we screen out input factors that have a significant impact on wind farm correction, while eliminating "noise" (input factors with low importance) that affect the correction results. Therefore, input factors with importance results greater than or equal to 0.05 are used as input variables for the subsequent multi-factor correction, while input factors with importance results less than or equal to 0.05 are eliminated as "noise."

[0027] From the calculation results of the characteristic importance of the input meteorological elements in the wind speed correction experiment, it can be seen that among the input meteorological elements, the importance calculation result of the 10m wind speed is the largest, and is much larger than the others. In addition to the 10m wind speed, the surface pressure (sp) and 2m dew point temperature (dt_2) also have relatively large importance calculation results.

[0028] The calculated characteristic importance of the input meteorological elements in the wind direction correction experiment shows that the 10-meter wind direction has the highest calculated importance, but its importance is relatively small compared to the wind speed. This is because wind direction errors are inherently random, resulting in large errors and making correction more difficult. In addition to the 10-meter wind direction, the importance of the 925hPa pressure layer wind direction (d_925), 10-meter wind speed (s10), 850hPa pressure layer wind direction (d_850), 850hPa pressure layer geopotential height (h_850), 2-meter dew point temperature (dt_2), and surface pressure (sp) also remained relatively high. In summary, analysis of the characteristic importance of wind speed and wind direction corrections determined the input meteorological elements for subsequent multi-factor input correction experiments. Among them, the wind speed correction selected three input elements: 10m wind speed, surface pressure and 2m dew point temperature; the wind direction correction selected seven input elements: 10m wind direction, 925hPa pressure layer wind direction, 10m wind speed, 850hPa pressure layer wind direction, 850hPa pressure layer potential height, 2m dew point temperature and surface pressure.

[0029] Step 6: Build an AU-Net convolutional neural network that integrates an attention mechanism, input the extracted training set into the neural network, and use the neural network's feedback system to continuously adjust the network structure to improve the model's accuracy in predicting other meteorological factors. To verify the improvement of the correction results with multi-factor input, when training the correction model, first only use the amount to be corrected as the model input and train a single-factor correction model. Then, based on the selected wind speed correction input meteorological factor, train a multi-factor correction model.

[0030] Based on the U-Net model, an attention mechanism is added to the jump connection to replace the original simple addition. The attention mechanism has been widely used in natural language processing, speech recognition, and image recognition. The idea of ​​the method is mainly derived from the human brain signal processing mechanism. Whether it is an image or a sentence, the human brain always pays attention to the important parts when processing information to increase efficiency and accuracy. The simplest understanding is to use weighting to represent the degree of importance. The AU-Net convolutional neural network structure that integrates the attention mechanism is shown in the attached figure. Figure 9 shown.

[0031] By implementing the attention module based on the soft-attention concept, the features of specific regions can be enhanced, and channel attention and spatial attention can be introduced into the model to further improve the model effect. The structure of the attention module is shown in the attached figure. Figure 10 Because wind speed correction is based on multiple factors, the most important of which is the numerically forecasted wind speed to be corrected, and other factors serve as supplements to the correction effect, a feature called channel attention is designed to have a greater impact on wind speed correction. Furthermore, considering spatial characteristics, a spatial attention mechanism is used to better learn the correlations between different locations, and to achieve better correction effects for outliers, such as areas with relatively high wind speeds.

[0032] The extracted training set is input into the constructed AU-Net neural network for training. The feedback system of the neural network is used to continuously adjust the network structure to improve the accuracy of the model's prediction of other meteorological factors. In order to verify the improvement of the correction results by multi-factor input, when training the correction model, only the amount to be corrected is used as the input of the model to train the single-factor correction model. Then, based on the selected wind speed correction input meteorological factors, including 10m wind speed, surface pressure and 2m dew point temperature, the multi-factor correction model is trained.

[0033] Step 7: Under the premise of single factor input, without considering the forecast timeliness, a verification evaluation is conducted based on the AU-Net model and the basic U-Net wind speed correction model, the prediction accuracy of the model is evaluated, and the results after model comparison are analyzed to improve the prediction accuracy of the regional medium- and long-term wind speed forecast correction model.

[0034] The correction model is the improved AU-Net. Before comparing the improvement of multi-factor input compared with single-factor input, we first conduct a wind speed correction accuracy evaluation experiment based on the AU-Net model and the basic U-Net model under the premise of single-factor input without considering the forecast timeliness. The experimental results are shown in Table 2.

[0035] Table 2. Model effect comparison table Model Name Model convergence mean loss (MSE) Test set RMSE Test set MAE U-Net 0.35 0.96 0.71 AU-Net 0.28 0.85 0.63 Whether it is the average loss of the final convergence during model training, or the root mean square error and mean absolute error of the corrected wind speed on the test set, the AU-Net model with the introduction of the attention mechanism is smaller than the UNet model, verifying the effectiveness of this method.

[0036] Attachment Figure 11 The wind speed errors of the original model, the random forest method (RF) correction, the single factor input method (AU-Net) correction, and the multi-factor input method (RF-AU-Net) correction are given. The left figure is the root mean square error, and the right figure is the mean absolute error. Figure 11 It can be seen that the effect of regional wind speed correction is significantly increased when only a single factor is input. This also shows that converting the numerical forecast correction problem into an image segmentation problem can make good use of the Encoder-Decoder model to extract errors and achieve good correction effects.

[0037] After introducing multiple variables, the root mean square error of different time periods was reduced to between 0.76-0.85m / s, with an average of about 0.80m / s, which is 62.8% higher than the numerical forecast results; the average absolute error of different time periods was between 0.56-0.62m / s, with an average of about 0.59m / s, which is 63.1% higher than the numerical forecast results. The effect was further improved, indicating that the introduction of relevant factors and the implementation of multi-factor input correction can further improve the effect of numerical forecast correction.

[0038] The wind speed error distribution has obvious seasonal characteristics. Figure 12 As can be seen from the figure, the amount of correction varies across months, with February achieving the best results. After correction, the root mean square error (RMS) dropped to 0.75 m / s, and the mean absolute error (MAE) decreased to 0.56 m / s. Months with average errors greater than the annual average (such as January, February, March, and December) showed more pronounced corrections than months with average errors less than the annual average (such as January, February, March, and December). This is because the principle of error backpropagation during training makes it easier to implement gradient descent for larger errors. Therefore, the model is more effective at correcting larger random errors. Seasonally, corrections were poor in spring and summer (April, May, June, and July), while they were better in autumn and winter (September, October, November, and December). Judging by the degree of reduction in error metrics, the multi-factor input correction method (RF-AU-Net) achieved excellent results for wind speed correction.

[0039] Compared to the random forest method, the corrected wind speed map based on the AU-Net network can better restore the actual wind speed distribution, especially for the larger wind speeds forecasted. This is due to the presence of jump connections. The wind speed can eliminate error information while retaining basic distribution information, resulting in a wind speed distribution closer to the label. Compared to the correction results of single-factor input, the wind speed correction results of multi-factor input retain more detailed information. Single-factor input will underestimate compared to multi-factor input, and when the wind speed is low, the effect is not as good as multi-factor correction, which is reflected in the wind speed image as a blurry image. In addition, the multi-factor correction method also has a good distinction between the boundary information of larger and smaller wind speeds. This shows that after introducing other relevant factors to participate in the correction, it can better distinguish the error distribution under different circumstances and retain more detailed information. This is particularly important for the regional forecast correction of wind speed.

Claims

1. A method for constructing a medium- to long-term wind speed prediction model for a single point, characterized by: The steps include: Step 1: Extract the training set of numerical forecast results. After filtering and organizing the data, extract the gridded numerical forecast products according to the n*n grid. The extracted parameters include temperature, pressure, humidity, wind field and precipitation. All this information needs to be integrated into the neural network for correction. Then, the corresponding grid forecast information is extracted from the historical near-day and near-time using the same method. At the same time, the information of ground observation data is also added to the training set to construct a high-quality training data set. Step 2: Using domain knowledge and feature engineering, select meteorological factors that have a significant impact on wind farm forecasts as input features, and perform data processing to remove irrelevant and noise features. Step 3: During data preprocessing, the grid data is expanded and treated as data from the same site, and a wind field correction model is constructed using the random forest algorithm. Step 4: Train and test the wind speed and direction correction models for different forecast timeframes, evaluate the correction effects for different forecast timeframes, and analyze the results. Step 5: Calculate the importance of the input factors for the numerical forecast wind field correction, select more representative correction input factors, and analyze the wind speed correction effect to achieve an improvement in the correction effect compared to single-factor input; Step 6: Build an AU-Net convolutional neural network with an integrated attention mechanism, input the extracted training set into the neural network, and use the neural network's feedback system to continuously adjust the network structure to improve the model's accuracy in predicting other meteorological factors. To verify that multi-factor input improves the correction results, when training the correction model, first use only the input to be corrected as the model input and train a single-factor correction model. Then, based on the selected wind speed correction input meteorological factor, train a multi-factor correction model. Step 7: Under the premise of single factor input, without considering the forecast timeliness, a verification evaluation is conducted based on the AU-Net model and the basic U-Net wind speed correction model, the prediction accuracy of the model is evaluated, and the results after model comparison are analyzed to improve the prediction accuracy of the regional medium- and long-term wind speed forecast correction model.

2. The method for constructing a single-point medium- and long-term wind speed prediction model according to claim 1, characterized in that: Based on the random forest algorithm, it includes the following steps: 3.

1. Assume that the size of the sample set is N, the number of decision trees in the algorithm is K, and the number of features of the sample is M; 3.

2. Construction of training sample set: Based on the original sample set, K new sample sets are randomly extracted using the bagging idea. Each sample set corresponds to a decision tree to be trained. 3.

3. Decision tree training: Utilize the idea of ​​feature subspace to perform the most reasonable attribute partitioning. That is, for M features, only m features are selected according to certain rules to participate in the most reasonable attribute partitioning, and a well-trained decision tree is obtained. 3.

4. Construction of random forest model: For k training sets, obtain K decision tree models according to the method in step 3.3, and obtain the wind field correction model based on random forest through ensemble learning.

3. The method for constructing a single-point medium- and long-term wind speed prediction model according to claim 1, characterized in that: Step 4 is to use the random forest method to calculate the importance of the input factors for the numerical forecast wind field correction and select more representative correction input factors. The main steps of the test are as follows: (1) Divide the training data according to the forecast time to obtain the wind speed and wind direction training data sets with different forecast time; (2) Build a wind field correction model based on random forest and train wind speed and wind direction correction models for different forecast time periods; (3) Test the effectiveness of the wind speed and direction correction model on the validation set, and adjust the training parameters, including the maximum depth of the tree and the number of trees, based on the validation set results; (4) A wind field correction test was conducted on the test set to evaluate the correction effects of different forecast timeliness and calculate the importance of input factors. The wind speed and wind direction correction models were both based on a regression model with mean square error as the loss. For the wind speed correction model, the number of trees was set to 50 and the maximum depth was set to 10. For the wind direction correction model, the number of trees was set to 60 and the maximum depth was set to 13.

4. The method for constructing a single-point medium- and long-term wind speed prediction model according to claim 1, characterized in that: Step five is as follows: for the calculation of feature importance, a purity index based on a decision tree is selected and normalized; for the importance calculation results based on random forest, input factors with importance results greater than or equal to 0.05 are used as input variables for subsequent multi-factor correction, and input factors with importance results less than or equal to 0.05 are removed as "noise".

Images