High plateau airport wind field prediction method based on numerical simulation and deep neural network

Abstract

The invention belongs to the technical field of meteorological prediction, and relates to a high plateau airport wind field prediction method based on numerical simulation and a deep neural network. The method comprises the following steps: performing mesoscale wind field numerical simulation by adopting a numerical weather forecast model, and outputting ground wind field prediction data of the high plateau airport; carrying out feature engineering processing; constructing a training data set to train the neural network model; verifying and testing the trained deep neural network model to obtain a high plateau airport wind field prediction model; and carrying out real-time high plateau airport wind field prediction. According to the high plateau region complex terrain, low-pressure and low-oxygen environment and local wind field characteristics, the numerical weather prediction model and the deep neural network are combined, multi-dimensional combined prediction is carried out on meteorological parameters, the wind field prediction precision is improved, and the short-term prediction precision of severe convective weather can be effectively improved; under the condition of lacking complete observation data, the accuracy of a prediction result can be ensured; and the method has relatively high universality and expandability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological prediction, and in particular, relates to a high-altitude airport wind field prediction method based on numerical simulation and deep neural networks. BACKGROUND

[0002] Due to the unique topography, climate and low-oxygen low-pressure environment of high-altitude areas, the wind field (including wind speed, wind direction and turbulence, etc.) has a particularly significant impact on flight take-off and landing safety, scheduling efficiency and aircraft operation. The complex topography (such as mountains and valleys) and severe local climate changes (such as strong convective weather and valley winds) in high-altitude areas make it difficult for traditional meteorological prediction methods to meet the requirements of high precision and real-time.

[0003] The prior art uses neural networks to correct the bias of numerical models, which performs well in global-scale weather prediction, but in high-altitude areas, the local effects caused by complex topography (such as valley winds and slope turbulence) are not fully considered, and the prediction accuracy in high-altitude areas is limited by resolution, physical parameterization schemes and sparsity of observation data.

[0004] The existing model training relies on a large amount of high-quality meteorological data or dense observation data, but the sparsity of observation sites in high-altitude areas leads to insufficient data, limiting the applicability of the method. Model training relies on, but the data sparsity in high-altitude areas limits the generalization ability of the model. In addition, the existing method is not combined with numerical models, and the prediction results may lack physical constraints, resulting in insufficient reliability under extreme weather conditions.

[0005] Airport wind field prediction needs to consider both large-scale weather systems (such as the westerlies and monsoons) and local wind field characteristics (such as microscale turbulence induced by topography), and the existing methods have obvious deficiencies in multi-scale coupling modeling. Existing meteorological prediction mainly targets low-altitude mountainous areas and does not fully consider the impact of low-pressure and low-oxygen environments on the atmospheric boundary layer in high-altitude areas, and lacks deep integration of numerical models and deep learning, resulting in a lack of physical consistency in the prediction results.

[0006] In addition, high-altitude airports have very high real-time requirements for wind field prediction (such as minute-level updates), but high-resolution numerical simulation and complex deep learning models such as GAN models (Generative Adversarial Networks) and LSTM (Long Short-Term Memory) have long computation times, and the existing methods are not optimized for airport minute-level real-time prediction requirements, resulting in long computation times and difficulty in meeting real-time requirements. SUMMARY

[0007] In order to solve the above technical problems, the application provides a high-altitude airport wind field prediction method based on numerical simulation and deep neural network, which comprises mesoscale numerical simulation, data feature engineering processing, deep neural network modeling, model verification and testing, and prediction and output. The mesoscale numerical simulation is performed by using a numerical weather prediction model to simulate the mesoscale wind field, and the high-altitude airport ground wind field prediction data is outputted, which comprises wind speed, wind direction, air pressure, temperature and humidity. The data feature engineering processing is performed on the high-altitude airport ground wind field prediction data, which comprises decomposing the wind speed and wind direction, collecting the meteorological feature data of the high-altitude region, and standardizing and normalizing all the obtained data; the meteorological feature data of the high-altitude region comprises time variable, air pressure, temperature, relative humidity, terrain height and terrain relief feature. The deep neural network modeling is performed by constructing a neural network model, using multiple sets of feature engineering processed high-altitude region meteorological feature data in a set time period containing several time steps as input data of the neural network model, using wind speed and wind direction as output data of the neural network model, and constructing a training data set to train the neural network model. The model verification and testing are performed by using cross-validation and actual observation data to verify and test the trained deep neural network model, evaluating the accuracy of the neural network model in wind field prediction, and obtaining a high-altitude airport wind field prediction model. The prediction and output are performed by using the high-altitude airport wind field prediction model to predict the real-time high-altitude airport wind field.

[0008] On the basis of the above technical solution, the application can also be improved as follows.

[0009] Further, the simulation process of the mesoscale numerical simulation adopts a three-layer one-way nested method for wind field numerical simulation, and sets the grid resolution and a plurality of vertical levels.

[0010] Further, the data assimilation method is used to fuse ground observation data, satellite remote sensing data and radar data, and the four-dimensional variation assimilation or ensemble Kalman filtering method is used to optimize the initial field and boundary conditions of the numerical weather prediction model.

[0011] Further, a combination of physical parameterization schemes is selected for simulation of the numerical weather prediction model; the physical parameterization schemes comprise planetary boundary layer scheme, microphysical scheme and radiation scheme; the numerical weather prediction model is used for 36-hour integral prediction, and the first 12 hours are the starting adjustment time, and the remaining 24 hours generate high-resolution wind field prediction data at intervals of 10 minutes.

[0012] Further, the data augmentation method is used to expand the training data set; the data augmentation method includes random disturbance and a generative adversarial network.

[0013] Further, the high-altitude area meteorological feature data after data feature engineering processing is subjected to principal component analysis or feature selection, so as to reduce the data dimension.

[0014] Further, the neural network model comprises an input layer, a hidden layer and an output layer; the hidden layer comprises a CNN layer, an LSTM layer and a full connection layer; the input data shape of the input layer is (60, 144, N), 60 represents the number of training sample groups, 144 represents 144 time steps of 24 hours, and N represents the number of features of each time step; a plurality of convolution kernels are used to extract spatial features in the output of the numerical weather prediction model through convolution operation; the spatial features include terrain-induced wind field change data; the activation function adopts a ReLU function; a plurality of LSTM units are used to capture the time sequence change features of the wind field; the full connection layer fuses the features extracted by the CNN layer and the LSTM layer, contains a plurality of neurons, and adds a Dropout layer; the output layer sets an output node, and the activation function adopts a linear activation function to output the output data of the neural network model.

[0015] Further, the high-altitude area meteorological feature data comprises time variables, air pressure, temperature, relative humidity, terrain height and terrain relief features and local meteorological features; the local meteorological features include valley wind data, slope turbulent flow data and wind shear feature data.

[0016] Further, the regularization terms of momentum conservation and energy conservation are added to the loss function to constrain the output of the neural network model.

[0017] Further, the error index between the wind field prediction result and the actual observation result of the wind field is obtained by calculating the mean square error and the mean absolute error, so as to obtain the evaluation result of the accuracy of the wind field prediction of the neural network model.

[0018] The beneficial effects of the present application are: (1) The present application is aimed at the complex terrain, low-pressure and low-oxygen environment and local wind field features of the high-altitude area, combines the physical modeling ability of the numerical weather prediction model and the self-adaptive learning ability of the deep neural network, improves the wind field prediction accuracy, and is suitable for high-altitude airports above 3000 meters and other airports greatly affected by terrain, and can effectively improve the short-term prediction accuracy of severe convective weather, and provide stronger protection for aviation flight safety; (2) The present application can perform rapid prediction calculation by using the neural network model, can better meet the efficient and rapid prediction demand of the complex airport environment of the high-altitude area, and can also ensure the accuracy of the prediction result in the case of lacking complete observation data; (3) The method of the present invention is not only applicable to wind field prediction at high-altitude airports, but can also be adjusted and expanded according to specific needs, and has strong universality and scalability. (4) In addition to wind field prediction, the method of this invention can also perform multi-dimensional joint prediction of meteorological parameters by combining numerical simulation results with deep neural networks. This provides technical support for the construction of a comprehensive meteorological early warning system. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the wind field prediction method for high-altitude airports based on numerical simulation and deep neural networks provided in Embodiment 1 of the present invention. Figure 2 A schematic diagram for selecting a combination of physical parameterization schemes. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] As an example, see the attached document. Figure 1 As shown, in order to solve the above-mentioned technical problems, this embodiment provides a method for predicting wind fields at high-altitude airports based on numerical simulation and deep neural networks, including mesoscale numerical simulation, data feature engineering processing, deep neural network modeling, model verification and testing, and prediction and output. Mesoscale numerical simulation: Centered on high-altitude airports, mesoscale wind field numerical simulation is performed using numerical weather prediction models, outputting predicted surface wind field data for high-altitude airports; the predicted surface wind field data for high-altitude airports includes wind speed, wind direction, air pressure, temperature and humidity; Data feature engineering processing: Feature engineering processing is performed on the surface wind field prediction data of high-altitude airports, including decomposing wind speed and direction, collecting meteorological feature data of high-altitude areas, and standardizing and normalizing all the obtained data; meteorological feature data of high-altitude areas include time variables, air pressure, temperature, relative humidity, terrain height and terrain relief characteristics. Deep Neural Network Modeling: Construct a neural network model by using multiple sets of feature-engineered meteorological feature data of high-altitude plateau regions with several time steps within a set time period as input data for the neural network model, and wind speed and wind direction as output data for the neural network model. Construct a training dataset to train the neural network model. Model validation and testing: The trained deep neural network model is validated and tested using cross-validation and actual observation data to evaluate the accuracy of the neural network model in wind field prediction and obtain a wind field prediction model for high-altitude airports. Prediction and Output: Real-time wind field prediction for high-altitude airports is performed using a high-altitude airport wind field prediction model.

[0022] This invention addresses the complex terrain, low-pressure and low-oxygen environment, and local wind field characteristics of high-altitude plateau regions. It combines the physical modeling capabilities of numerical weather prediction models with the adaptive learning capabilities of deep neural networks to improve wind field prediction accuracy. This invention is applicable to high-altitude airports above 3000 meters and other airports significantly affected by terrain, effectively improving the short-term prediction accuracy of severe convective weather and providing stronger protection for aviation safety.

[0023] Traditional numerical weather prediction models have limited accuracy in wind field prediction in high-altitude plateau regions, mainly due to the imperfect simulation of physical processes under complex terrain and climate conditions. This invention introduces a neural network model into the prediction process, which can fully leverage the adaptability of deep learning and capture nonlinear relationships that are difficult to express by traditional numerical simulation methods, thereby significantly improving the accuracy of wind field prediction.

[0024] Numerical simulation provides deep neural networks with rich meteorological data features, which in turn enable deep neural networks to learn complex patterns and laws from these data, thereby improving the accuracy and robustness of predictions while ensuring physical plausibility.

[0025] Wind fields in high-altitude plateau regions are subject to complex influences from various factors such as topography, airflow patterns, temperature, and humidity. Traditional numerical simulation methods struggle to address these complex meteorological phenomena. To adapt to the unique climate of high-altitude plateau regions, this invention utilizes a deep neural network model to learn the climate characteristics of these regions from historical meteorological data. By automatically adjusting the parameters of the neural network model to adapt to these special climatic conditions, it achieves high-precision wind field prediction.

[0026] Neural network models can perform rapid prediction calculations, especially after training and optimization, enabling them to quickly output prediction results after real-time data input. This is of great significance for real-time wind field monitoring and flight safety at high-altitude airports. Compared to traditional numerical simulation methods that require long computation times, the method of this invention can better meet the demand for efficient and rapid prediction in the complex environment of high-altitude airports.

[0027] Adaptive optimization and flexibility: Through training and optimization of deep neural networks, the model can be flexibly applied and adjusted for different high-altitude airports. Even in the absence of complete observation data, deep neural networks can adaptively adjust through their powerful learning capabilities, thereby ensuring the accuracy of prediction results.

[0028] The method of this invention is not only applicable to wind field prediction at high-altitude airports, but can also be adjusted and extended according to specific needs to be applied to wind field prediction under other complex meteorological conditions. By changing the input features and adjusting the network structure, it can be extended to wind field prediction scenarios in other regions, and has strong universality and scalability.

[0029] In addition to wind field prediction, the method of this invention can also perform multi-dimensional joint prediction of meteorological parameters such as temperature, humidity, and air pressure by combining numerical simulation results with deep neural networks. This provides technical support for the construction of a comprehensive meteorological early warning system.

[0030] Deep neural networks can continuously optimize their prediction models through long-term training and use. With the input of more historical and real-time data, the predictive ability of neural network models will gradually improve, further enhancing the reliability and stability of long-term predictions.

[0031] Optionally, the mesoscale numerical simulation process employs a three-layer unidirectional nested method for wind field numerical simulation, setting the grid resolution and several vertical layers.

[0032] In practical applications, the simulation area is centered on a high-altitude airport, and the WRF (Weather Research and Forecasting Model) is used for mesoscale numerical simulation. The simulation area uses a three-layer unidirectional nested approach, with grid resolutions of 8000 meters, 2000 meters, and 500 meters for the outer, middle, and inner layers, respectively. The inner layer covers the airport terminal area, focusing on the refined prediction of near-surface wind fields, effectively capturing local wind field characteristics (such as valley winds and slope turbulence) caused by complex terrain (such as mountains and canyons). The simulation area is set with 65 vertical layers, and the resolution of the near-ground boundary layer is specially optimized. For example, the vertical coordinate parameters of the first 10 layers are set to 0.996, 0.988, 0.978, 0.966, 0.956, 0.946, 0.933, 0.923, 0.912 and 0.901, respectively, to ensure that there is sufficient vertical resolution within the boundary layer to accurately simulate turbulence and thermodynamic processes in the low-pressure and low-oxygen environment of high-altitude plateau regions.

[0033] Optionally, a data assimilation method can be used to fuse ground observation data, satellite remote sensing data, and radar data, and the initial field and boundary conditions of the numerical weather prediction model can be optimized through four-dimensional variational assimilation or ensemble Kalman filtering.

[0034] To improve the initial field accuracy of the WRF model, a data assimilation method is employed to integrate ground observation data, including automatic airport observation data (such as wind speed, wind direction, temperature, air pressure, and humidity). The initial field and boundary conditions of the WRF model are optimized using four-dimensional variational assimilation or ensemble Kalman filtering methods to reduce the uncertainty caused by the sparsity of observation data in high-altitude plateau regions.

[0035] Optionally, a combination of physical parameterization schemes can be selected for numerical weather prediction model simulation; physical parameterization schemes include planetary boundary layer scheme, microphysics scheme and radiation scheme; the numerical weather prediction model is used for 36-hour integral prediction, with the first 12 hours as the start-up adjustment time, and the remaining 24 hours generating high-resolution wind field prediction data at 10-minute intervals.

[0036] In response to the unique atmospheric environment of high-altitude plateau regions, as shown in the attached document... Figure 2 As shown, a preferred combination of physical parameterization schemes is included, such as the planetary boundary layer scheme, the microphysics scheme, and the radiation scheme. Figure 1As shown. Planetary boundary layer schemes include: ACM2 (Advanced Canopy Model 2, simulating energy and mass exchange between the vegetation canopy and the atmospheric boundary layer), MYJ (Mellor-Yamada-Janjic scheme, simulating turbulent exchange within the boundary layer), YSU (Yonsei University scheme, considering boundary layer development under different thermodynamic and dynamic conditions), and MYNN2 (Mellor-Yamada-Nakanishi-Niino 2 scheme, optimizing the simulation of boundary layer turbulent mixing processes); microphysics schemes include MP2 (Simple ice (2-class) scheme, simulating microphysics processes such as clouds and precipitation), MP16 (More complex microphysics with 16 categories, refining the classification of water-forming substances (such as various ice crystals, water droplets, and graupel), simulating cloud and precipitation microphysics processes), and MP22 (Further refined microphysics with 22... Categories further refine the classification of aquatic organisms and improve the accuracy of microphysical process simulations) and MP50 (High-resolution microphysics with 50+ categories, a high-resolution microphysics scheme that finely classifies aquatic organisms for high-precision simulation needs); radiation schemes include: GODDARD (Goddard Radiation Scheme, which simulates atmospheric radiative transfer processes such as solar shortwave radiation, atmospheric longwave radiation, and their interaction with the atmosphere and the Earth's surface), RRTM (Rapid Radiative Transfer Model, which efficiently calculates atmospheric longwave radiative transfer and balances accuracy and computational efficiency in numerical simulations) and RRTMG (Rapid Radiative Transfer Model for GCMs, a rapid radiative transfer model applicable to global climate models).

[0037] Optionally, data augmentation methods can be used to expand the training dataset; data augmentation methods include random perturbation and generative adversarial networks.

[0038] Optionally, principal component analysis or feature selection can be performed on the meteorological characteristic data of high-altitude plateau regions after data feature engineering to reduce data dimensionality.

[0039] Optionally, the neural network model includes an input layer, hidden layers, and an output layer; the hidden layer includes a CNN layer, an LSTM layer, and a fully connected layer; the input data shape of the input layer is set to (60, 144, N), where 60 represents the number of training sample groups, 144 represents the 144 time steps in 24 hours, and N represents the number of features at each time step; several convolutional kernels are used to extract spatial features from the output of the numerical weather prediction model through convolution operations; spatial features include wind field change data caused by terrain; the activation function is the ReLU function; several LSTM units are used to capture the time series change features of the wind field; the fully connected layer fuses the features extracted by the CNN layer and the LSTM layer, contains several neurons, and a Dropout layer is added; the output layer sets the output node, uses a linear activation function, and outputs the output data of the neural network model.

[0040] As an optional implementation, the CNN layer uses 128-512 convolutional kernels (the specific number is adjusted according to the task complexity) to extract spatial features (such as wind field changes caused by terrain) from the WRF model output through convolutional operations. The activation function adopts ReLU (Rectified Linear Unit) to enhance non-linear expressive power; the LSTM layer uses 64-256 LSTM units to capture the time-series evolution of the wind field and adapt to the dynamic changes of the wind field in high-altitude areas; the fully connected layer fuses the features extracted by the CNN layer and the LSTM layer, containing 64-128 neurons, and adds a Dropout layer (dropout rate 0.2-0.5) to prevent overfitting; for the regression task of wind field prediction, the output layer is designed with 2 nodes (outputting 2 wind direction components respectively) or 1 node (outputting wind speed or wind direction), and the activation function is a linear activation function to retain continuous value output.

[0041] Optional meteorological characteristic data for high-altitude plateau regions include time variables, air pressure, temperature, relative humidity, topographic altitude and topographic relief characteristics, and local meteorological characteristics; local meteorological characteristics include valley wind data, slope turbulence data, and wind shear characteristic data.

[0042] Optionally, regularization terms for momentum and energy conservation can be added to the loss function to constrain the output of the neural network model.

[0043] By incorporating regularization terms based on momentum and energy conservation into the loss function, the physical consistency of the model output is constrained, thus addressing the non-physical results that may occur in traditional deep learning models under extreme weather conditions.

[0044] Optionally, the error index between the wind field prediction result and the actual wind field observation result can be obtained by calculating the mean square error and the mean absolute error, which can be used as an evaluation result of the accuracy of the wind field prediction of the neural network model.

[0045] During the training of the neural network model, historical meteorological data and observational data are used to train the deep neural network. The training process optimizes the network weights through the backpropagation algorithm to minimize the prediction error. For example, if the mean squared error (MSE) is used as the loss function, the optimization objective is to minimize the error between the predicted wind field and the actual observed wind field.

[0046] Optionally, the Adam optimizer (with an initial learning rate of 0.001) can be used for model training, combined with an early stopping mechanism to prevent overfitting. The neural network model can be validated using cross-validation methods (e.g., K-fold cross-validation, K=5). The feature dataset is divided into training and validation sets to ensure the robustness of the neural network model under different weather conditions. The 60 data sets are divided into training, validation, and test sets (e.g., with proportions set to 80%, 10%, and 10% respectively).

[0047] As an alternative implementation, rolling wind field predictions can be achieved every 10 minutes through GPU parallel computing and model compression methods (such as pruning and quantization) to meet the real-time requirements of high-altitude airports.

[0048] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A high-altitude airport wind field prediction method based on numerical simulation and deep neural network, characterized in that, The method comprises the following steps: a. mesoscale numerical simulation, b. data feature engineering processing, c. deep neural network modeling, d. model verification and testing, and e. prediction and output. The mesoscale numerical simulation is performed by using a numerical weather prediction model, and the high-altitude airport ground wind field prediction data including wind speed, wind direction, air pressure, temperature and humidity are outputted. The data feature engineering processing is performed on the high-altitude airport ground wind field prediction data, including decomposing the wind speed and wind direction, collecting the high-altitude area meteorological feature data, and standardizing and normalizing all the obtained data. The deep neural network modeling is performed by constructing a neural network model, using the high-altitude area meteorological feature data processed by the feature engineering as the input data of the neural network model, and using the wind speed and wind direction as the output data of the neural network model. The model verification and testing are performed by using the cross-validation and actual observation data to verify and test the trained deep neural network model, evaluating the accuracy of the wind field prediction of the neural network model, and obtaining the high-altitude airport wind field prediction model. The prediction and output are performed by using the high-altitude airport wind field prediction model to perform real-time high-altitude airport wind field prediction.

2. The high-altitude airport wind field prediction method based on numerical simulation and deep neural network according to claim 1, characterized in that, The simulation process of the mesoscale numerical simulation is performed by using a three-layer one-way nested method for wind field numerical simulation, and the grid resolution and a plurality of vertical levels are set.

3. The method of claim 1, wherein, The data assimilation method is used to fuse the ground observation data, satellite remote sensing data and radar data, and the four-dimensional variational assimilation or ensemble Kalman filter method is used to optimize the initial field and boundary conditions of the numerical weather prediction model.

4. The high-altitude airport wind field prediction method based on numerical simulation and deep neural network according to claim 1, characterized in that, The combination of physical parameterization schemes is selected for the simulation of the numerical weather prediction model; the physical parameterization schemes include planetary boundary layer scheme, microphysical scheme and radiation scheme; the numerical weather prediction model is used for 36-hour integral prediction, and the remaining 24 hours are generated at an interval of 10 minutes to generate high-resolution wind field prediction data.

5. The method of claim 1, wherein, The data augmentation method is used to expand the training data set; the data augmentation method includes random disturbance and generative adversarial network.

6. The method of claim 1, wherein, The high-altitude area meteorological feature data processed by the data feature engineering is subjected to principal component analysis or feature selection to reduce the data dimension.

7. The method of claim 1, wherein, The neural network model comprises an input layer, a hidden layer and an output layer; the hidden layer comprises a CNN layer, an LSTM layer and a full connection layer; the input data shape of the input layer is (60, 144, N), 60 represents the number of training sample groups, 144 represents 144 time steps of 24 hours, and N represents the number of features of each time step; spatial features in the output of the numerical weather prediction model are extracted by using a plurality of convolution kernels through a convolution operation; the spatial features comprise terrain-induced wind field change data; the activation function adopts a ReLU function; a plurality of LSTM units are used to capture the time sequence change features of the wind field; the full connection layer fuses the features extracted by the CNN layer and the LSTM layer, contains a plurality of neurons, and adds a Dropout layer; the output layer sets an output node, the activation function adopts a linear activation function, and the output data of the neural network model is output.

8. The method of claim 1, wherein, The meteorological feature data of the high-altitude plateau region comprises time variables, air pressure, temperature, relative humidity, terrain height, terrain relief features and local meteorological features; the local meteorological features comprise valley wind data, slope turbulent flow data and wind shear feature data.

9. The method of claim 1, wherein, The regularization terms of momentum conservation and energy conservation are added to the loss function to constrain the output of the neural network model.

10. The method of claim 1, wherein, The error index between the wind field prediction result and the actual observation result of the wind field is obtained by calculating the mean square error and the mean absolute error, and is used as the evaluation result of the accuracy of the wind field prediction of the neural network model.

Images