An airport visibility prediction method and device based on machine learning, an electronic device, and a storage medium

Abstract

The application relates to the technical field of aviation meteorological prediction, in particular to an airport visibility prediction method and device based on machine learning, electronic equipment and a storage medium, which fuses ECMWF numerical prediction, DEM terrain elevation and CAMS air pollution multi-source data, constructs combined cross features guided by physical mechanism, and enriches the input information dimension; a residual learning strategy is adopted to learn only the ECMWF prediction residual, so that the learning difficulty is reduced and the physical basis is reserved; a multi-scale hollow convolution is used to extract a multi-scale time sequence mode, day and night and season priori are combined with physical perception position coding to be injected, and a terrain-meteorological self-adaptive modulation module uses terrain features and global time sequence context to generate modulation parameters, dynamically modulates the encoder output, realizes self-adaptive adjustment of terrain influence with weather background, improves the airport multi-time visibility prediction precision and generalization ability, and meets the aviation runway level refined prediction demand.

Description

Technical Field

[0001] This invention relates to the field of aviation weather forecasting technology, and in particular to a machine learning-based method, device, electronic equipment, and storage medium for predicting airport visibility. Background Technology

[0002] Visibility is a core meteorological element measuring atmospheric transparency, and low visibility weather (such as dense fog, haze, and heavy precipitation) is one of the key factors affecting aviation safety and operational efficiency. During airport takeoff and landing, visibility directly determines runway visual range and landing standard rating; low visibility can lead to widespread flight delays, diversions, and even accidents. Therefore, accurate airport visibility forecasts are of great value for air traffic control, flight planning, and passenger safety.

[0003] Currently, visibility forecasting mainly relies on the following three types of methods: (1) Direct forecasting methods based on numerical weather prediction models, such as ECMWF numerical forecast products, which have been widely adopted by meteorological departments around the world. However, numerical models inevitably have partial differential equation approximation errors when describing atmospheric motion and heat exchange, and there are always systematic and random errors, especially in low visibility forecasts under complex terrain and pollution conditions, where the errors are more significant; (2) Forecast correction methods based on traditional statistics, mainly including model output statistical methods, Kalman filtering and multiple linear regression, etc. However, these methods rely heavily on the forecaster's experience and subjective judgment. The correction methods are mainly aimed at individual meteorological elements, ignoring the correlation characteristics between meteorological elements, and the ability to make single-point fine forecasts for complex terrain areas is limited; (3) Visibility prediction methods based on machine learning. In recent years, random forests, support vector machines, XGBoost, LightGBM and various deep learning networks have been widely used in meteorological element forecast correction, which can learn the complex nonlinear relationship between the forecast field and the observation field.

[0004] However, existing machine learning-based visibility prediction methods have failed to effectively integrate multi-source heterogeneous data and make full use of the physical constraint information contained therein, making it difficult to meet the actual needs of aviation operations for runway-level refined visibility forecasts. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a machine learning-based airport visibility prediction method, device, electronic device, and storage medium.

[0006] In a first aspect, embodiments of the present invention provide an airport visibility prediction method based on machine learning, the method comprising: Acquire multi-source heterogeneous data, which should include at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data; Multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing value processing and standardization preprocessing. The preprocessed multi-source data is spliced ​​and fused to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms. The feature vector set is input into a pre-trained meteorological adaptive residual correction network, which outputs the visibility prediction results of the target airport in the future multiple time periods. The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected sequentially. Based on the ECMWF visibility forecast, the network extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and then uses the terrain features extracted from the DEM terrain elevation data and the global context vector output by the Transformer encoder to generate modulation parameters, dynamically modulating the encoder output sequence. The decoding output layer then decodes and outputs the visibility residual prediction values ​​for each time period. These residual prediction values ​​are added to the original ECMWF visibility forecast values ​​to obtain the corrected visibility prediction values, which serve as the visibility prediction results.

[0007] In conjunction with the first aspect, the steps for preprocessing multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data include: ECMWF numerical forecast data are interpolated to the latitude and longitude coordinates of the target airport using the bilinear interpolation method to obtain time series data of ECMWF meteorological elements matched to the target airport; The CAMS air pollution forecast data is interpolated to the latitude and longitude coordinates of the target airport using the bilinear interpolation method to obtain the time series data of each pollution element of CAMS matched to the target airport; The elevation values ​​of the target airport runway location are extracted from the DEM topographic elevation data, and the slope characteristics and topographic relief characteristics are calculated based on the DEM grid points of the surrounding area of ​​the target airport to obtain the DEM topographic elevation data of the target airport. The airport's measured visibility label data is matched and aligned precisely with the interpolated ECMWF time series data, CAMS time series data, and DEM topographic elevation data according to the timestamp, to obtain a time-aligned multi-source fusion sample dataset. Missing values ​​are detected in the multi-source fusion sample dataset, and time samples with missing values ​​are deleted to obtain a cleaned complete sample dataset. Z-score standardization is performed on all continuous numerical variables in the complete sample dataset to convert each variable to a uniform scale, resulting in preprocessed multi-source heterogeneous data.

[0008] In conjunction with the first aspect, the steps for feature splicing and fusion of preprocessed multi-source data include: The basic feature vector is obtained by stitching together the meteorological elements in the preprocessed ECMWF numerical forecast data, the pollution elements in the CAMS air pollution forecast data, and the DEM topographic elevation data. Based on meteorological elements in continuous time intervals from ECMWF numerical forecast data, the trend values ​​and statistics of each meteorological element within the preset sliding time window are calculated to obtain the statistical characteristics of the time-series sliding window. Based on the physical mechanism of atmospheric extinction, derived features of cross-combination are extracted from ECMWF numerical forecast data and CAMS air pollution forecast data to obtain combined cross-features based on physical mechanisms; The basic feature vector, time-series sliding window statistical features, and combined cross features are concatenated to form a complete feature vector set.

[0009] In conjunction with the first aspect, the steps of inputting the feature vector set into a pre-trained meteorological adaptive residual correction network and outputting the visibility prediction results for the target airport in multiple future timeframes include: The feature vector set is organized in chronological order into a mixed time series containing historical data segments and future forecast segments; The hybrid time series is input into the meteorological adaptive residual correction network. The local time series patterns at different time scales are extracted by multiple parallel one-dimensional dilated convolutions with different dilation rates. The outputs of each parallel dilated convolution are then concatenated to obtain multi-scale time series features. Multi-scale temporal features are added element by element to the physical sensing location code to obtain temporal features that fuse location information; The temporal features of the fused location information are input into the Transformer encoder, and processed by a multi-head self-attention mechanism and a feedforward network to capture the long-range dependencies between different time points in the hybrid temporal sequence, thus obtaining the encoder feature sequence. The encoder feature sequence is input into the terrain-meteorological adaptive feature modulation module. Modulation parameters are generated based on the global temporal context of DEM terrain elevation data and encoder feature sequence. The encoder feature sequence is dynamically modulated using the modulation parameters to obtain the terrain-meteorological adaptive modulated feature sequence. The time steps corresponding to the future forecast segments are extracted from the modulated feature sequences and input into the decoding output layer for mapping to obtain the visibility residual prediction values ​​for each time period. Obtain the original visibility forecast values ​​for each forecast time period from the ECMWF numerical weather prediction data. Add the visibility residual prediction values ​​for each time period to the original visibility forecast values ​​for each time period to obtain the corrected visibility prediction values ​​for each time period. Use these corrected visibility prediction values ​​for the target airport for multiple time periods in the future and output them.

[0010] In conjunction with the first aspect, before the step of adding multi-scale temporal features and physical sensing location codes element-wise to obtain temporal features with fused location information, the following steps are also included: Obtain the absolute position of each time step in the mixed time series sequence, and generate a relative position encoding vector through sinusoidal position encoding; Obtain the real-world time corresponding to each time step, calculate the solar hour angle based on the real-world time to represent day and night information, and calculate the annual day to represent seasonal information; Sine coding is performed on the solar hour angle and the annual day, and the coding results are mapped to the hidden dimension through a fully connected layer to obtain the physical time coding vector. The relative position encoding vector is added element by element to the physical time encoding vector to obtain the physical sensing position encoding.

[0011] In conjunction with the first aspect, the steps of dynamically modulating the encoder feature sequence using modulation parameters to obtain the terrain-weather adaptively modulated feature sequence include: Input the DEM topographic elevation data into the topographic-meteorological adaptive feature modulation module to obtain a static topographic feature vector; The encoder feature sequence is processed by global average pooling to obtain the global temporal context vector; The static terrain feature vector is concatenated with the global temporal context vector to obtain the fused feature vector; The fused feature vectors are input into a fully connected network for mapping, and the outputs a scaling parameter sequence and a shift parameter sequence. The scaling parameter sequence and the shift parameter sequence are applied to each feature channel of the encoder feature sequence respectively, and the encoder feature sequence is modulated channel by channel to obtain the feature sequence after terrain-weather adaptive modulation.

[0012] In conjunction with the first aspect, the steps of extracting the time steps corresponding to the future forecast segments from the modulated feature sequence, inputting them into the decoding output layer for mapping, and obtaining the visibility residual prediction values ​​for each time period include: From the feature sequence after terrain-meteorological adaptive modulation, feature vectors corresponding to each forecast lead time of the future forecast segment are extracted; The feature vectors of each forecast time are extracted and input into the decoding output layer. After passing through two fully connected networks, the visibility residual prediction value corresponding to each forecast time is output.

[0013] Secondly, this application provides an airport visibility prediction device based on machine learning, the device comprising: The acquisition module is used to acquire multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data. The preprocessing module is used to preprocess multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing value processing and standardization preprocessing; The fusion module is used to perform feature splicing and fusion on the preprocessed multi-source data to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms. The prediction module is used to input the feature vector set into a pre-trained meteorological adaptive residual correction network and output the visibility prediction results of the target airport in the future multiple time periods. The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected sequentially. Based on the ECMWF visibility forecast, the network extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and then uses the terrain features extracted from the DEM terrain elevation data and the global context vector output by the Transformer encoder to generate modulation parameters, dynamically modulating the encoder output sequence. The decoding output layer then decodes and outputs the visibility residual prediction values ​​for each time period. These residual prediction values ​​are added to the original ECMWF visibility forecast values ​​to obtain the corrected visibility prediction values, which serve as the visibility prediction results.

[0014] Thirdly, this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor runs the computer program to cause the electronic device to perform the above-described method.

[0015] Fourthly, this application provides a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0016] The embodiments of the present invention bring the following beneficial effects: This application provides a machine learning-based airport visibility prediction method, device, electronic device, and storage medium. The method includes: acquiring multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data; preprocessing the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing measurement processing, and standardization preprocessing; performing feature splicing and fusion on the preprocessed multi-source data to construct a feature vector set containing basic features, temporal sliding window statistical features, and combined cross features based on physical mechanisms; inputting the feature vector set into a pre-trained meteorological adaptive residual correction network to output the visibility prediction results of the target airport in multiple future time periods; wherein, the meteorological adaptive residual correction network includes a multi-scale dilated convolutional layer, a physical sensing position encoding layer, a Transformer encoder, a topographic-meteorological adaptive feature modulation module, and a decoding output layer connected in sequence.

[0017] This application enriches the information dimensions of visibility prediction by fusing multi-source data from ECMWF numerical weather prediction, DEM topographic elevation, and CAMS air pollution, and constructing a combined cross-feature based on physical mechanisms. It employs a residual learning strategy, learning only the residual portion of the ECMWF visibility forecast, reducing learning difficulty while preserving the physical basis of the numerical model. Through multi-scale dilated convolution, it extracts temporal patterns at different time scales, injects prior information on day / night and seasons by combining physical sensing location coding, and uses a terrain-meteorological adaptive modulation module to generate modulation parameters based on terrain features and global temporal context to dynamically modulate the encoder output sequence. This achieves adaptive adjustment of terrain influence with weather background, improving the accuracy and generalization ability of single-point, multi-time-effect visibility prediction for airports, effectively meeting the actual needs of aviation operations for runway-level refined visibility forecasting.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a machine learning-based airport visibility prediction method provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the principle of the method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the principle of the terrain-meteorological adaptive feature modulation module in the method provided in this embodiment of the invention; Figure 4 A schematic diagram of an airport visibility prediction device based on machine learning provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention.

[0022] Figure label: 10 - Acquisition module, 20 - Preprocessing module, 30 - Fusion module, 40 - Prediction module; 130 - Processor, 131 - Memory, 132 - Bus, 133 - Communication interface. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To facilitate understanding of this embodiment, the technical terms used in this application will be briefly introduced below.

[0025] The Transformer encoder is a deep learning network structure based on the self-attention mechanism. It uses multi-head self-attention to capture long-range dependencies between any two positions in a sequence, solving the problem that recurrent neural networks and convolutional neural networks are difficult to compute in parallel and capture long-range dependencies when processing long sequences.

[0026] The residual learning strategy refers to a network that does not directly learn the absolute value of the target variable, but instead learns the residual (i.e., the difference) between the target variable and the existing baseline value.

[0027] Dilated convolution expands the receptive field by inserting holes between convolutional kernel elements, enabling the network to capture a wider range of contextual information without increasing the number of parameters. Multi-scale dilated convolution uses multiple parallel dilated convolutions with different dilation rates to extract feature patterns across different time spans.

[0028] After introducing the technical terms used in this application, the application scenarios and design concepts of the embodiments of this application will be briefly described below.

[0029] Existing machine learning visibility prediction methods do not effectively integrate terrain elevation and air pollution information, lack guidance from atmospheric extinction physical mechanisms, and are mostly targeted at regional grid forecasts. They also lack refined correction schemes for single-point, multi-time-based airport forecasts, resulting in insufficient accuracy in low visibility forecasts.

[0030] Based on this, embodiments of this application provide a machine learning-based airport visibility prediction method, device, electronic device, and storage medium.

[0031] Example 1 This application provides a machine learning-based airport visibility prediction method, combined with... Figure 1 As shown, the method includes: S110. Acquire multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data.

[0032] S120, preprocess the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing measurement value processing and standardization preprocessing.

[0033] S130 performs feature splicing and fusion on the preprocessed multi-source data to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms.

[0034] S140 inputs the feature vector set into the pre-trained meteorological adaptive residual correction network and outputs the visibility prediction results of the target airport in the future multiple time periods.

[0035] The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected sequentially. Based on the ECMWF visibility forecast, the network extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and then uses terrain features extracted from the DEM terrain elevation data (combined with...) by the terrain-meteorological adaptive feature modulation module. Figure 2 The modulation parameters are generated from the static features of the DEM and the global context vector output by the Transformer encoder. The encoder output sequence is dynamically modulated, and the visibility residual prediction values ​​for each time period are decoded by the decoding output layer. The residual prediction values ​​are added to the original visibility forecast values ​​of ECMWF to obtain the corrected visibility prediction values, which are used as the visibility prediction results.

[0036] This application constructs a combined cross-feature based on physical mechanisms by fusing multi-source heterogeneous data from ECMWF numerical weather prediction, DEM topographic elevation, and CAMS air pollution data, thus enriching the information input dimensions for visibility prediction. It employs a residual learning strategy, learning only the residual portion of the ECMWF visibility forecast, reducing the model learning difficulty while preserving the physical basis of the numerical model. Furthermore, it extracts temporal patterns at different time scales through multi-scale dilated convolution, injects prior information on day / night and seasons by combining physical sensing location coding, and utilizes a terrain-meteorological adaptive modulation module to dynamically adjust the terrain influence according to the weather background, thereby improving the accuracy and generalization ability of single-point, multi-time-effect visibility prediction for airports.

[0037] In step S110, the ECMWF numerical weather prediction data refers to the global numerical weather prediction products published by the European Centre for Medium-Range Weather Forecasts (ECMWF). In this embodiment, historical reanalysis forecast data and real-time forecast data of the target airport location are obtained through the ECMWF data interface. This data includes at least gridded forecast data of the following meteorological elements: 2-meter temperature, 2-meter dew point temperature, 10-meter wind speed (U / V component), sea level pressure, relative humidity, cloud cover, total precipitation, and ECMWF visibility forecast values. Taking a certain airport as an example, the acquired ECMWF data is gridded data every 3 hours, covering a forecast lead time of 0-72 hours, with a spatial resolution of 0.1°×0.1°, and a data time range from January 2021 to December 2025.

[0038] DEM topographic elevation data refers to the surface elevation information recorded by a Digital Elevation Model. In this embodiment, high-resolution topographic data of the target airport area is obtained through public data sources (such as SRTM or CopernicusDEM). This data includes at least the elevation values ​​of the target airport runway location, as well as topographic statistics calculated based on the surrounding area of ​​the airport (such as a 3×3 or 5×5 grid window around the runway center point), including slope characteristics and topographic relief characteristics. Topography has a significant impact on local atmospheric flow and temperature and humidity distribution, especially under complex terrain conditions. Introducing DEM data can effectively characterize the constraint effect of local topography on visibility changes.

[0039] CAMS air pollution forecast data refers to the global atmospheric composition analysis and forecast data released by the Copernicus Atmosphere Monitoring Service. In this embodiment, near-real-time analysis and forecast data of global atmospheric composition at the target airport location are obtained through the CAMS data interface. This data includes gridded forecast data for at least the following pollutants: PM2.5 concentration, PM10 concentration, aerosol optical depth (AOD, 550nm), nitrogen dioxide concentration, ozone concentration, and sulfate aerosol concentration, with a temporal resolution of 3 hours and a forecast lead time of 0-72 hours. Taking a certain airport as an example, the acquired CAMS data is gridded data with a 3-hour time interval and a forecast lead time of 0-72 hours, with a spatial resolution of 0.4°×0.4°. Aerosol particulate matter in air pollutants is an important factor affecting the atmospheric extinction coefficient, especially during haze and fog weather, where it directly affects visibility. Introducing CAMS data can enhance the model's ability to perceive atmospheric extinction effects.

[0040] Airport measured visibility label data refers to the actual visibility values ​​recorded by the target airport's automatic weather observation system or manual observation. In this embodiment, hourly measured visibility data is obtained through the target airport's Aviation Weather Report (METAR) or Automated Weather Observation System (AWOS). This data serves as the ground truth label for model training, used to calculate residual training objectives and verify model prediction accuracy. Referring to the above example, the METAR message provides dominant visibility observations at hourly or half-hourly intervals, ranging from 0 to 10,000 meters. When visibility is below 10,000 meters, a specific value is recorded; when visibility is above 10,000 meters, it is recorded as 9999 or CAVOK. The measured visibility data is matched and aligned with the above three types of data according to the timestamp, ensuring that the forecast data at each moment can establish a mapping relationship with the corresponding measured label, providing a reliable ground truth basis for model training and accuracy verification.

[0041] After acquiring the above four types of data, they need to be matched and aligned according to a unified timeline to ensure that ECMWF forecast data, CAMS pollution data, DEM topographic features, and measured visibility labels at the same time can be correlated, providing a data foundation for subsequent preprocessing and feature engineering. By integrating four heterogeneous data sources—numerical forecasts, topographic elevations, air pollution, and measured labels—this embodiment can characterize the key factors affecting visibility from multiple dimensions of meteorology, topography, and pollution, providing rich information input for deep learning models.

[0042] In conjunction with the first aspect, step S120 includes: S121. The ECMWF numerical forecast data is interpolated to the latitude and longitude coordinates of the target airport using the bilinear interpolation method to obtain the time series data of each meteorological element of ECMWF matched to the target airport.

[0043] S122, the CAMS air pollution forecast data is interpolated to the latitude and longitude coordinates of the target airport using the bilinear interpolation method to obtain the time series data of each pollution element of CAMS matched to the target airport.

[0044] S123: Extract the elevation value of the target airport runway location from the DEM topographic elevation data, and calculate the slope characteristics and topographic relief characteristics based on the DEM grid points of the surrounding area of ​​the target airport to obtain the DEM topographic elevation data of the target airport.

[0045] S124. The airport's measured visibility label data is matched and aligned precisely with the interpolated ECMWF time series data, CAMS time series data, and DEM topographic elevation data according to the timestamp, to obtain a time-aligned multi-source fusion sample dataset.

[0046] S125, perform missing value detection on the multi-source fusion sample dataset, delete time samples with missing values, and obtain the cleaned complete sample dataset.

[0047] S126. Perform Z-score standardization on all continuous numerical variables in the complete sample dataset to convert each variable to a uniform scale, resulting in preprocessed multi-source heterogeneous data.

[0048] In step S121, the ECMWF grid forecast data obtained in step S110 is interpolated to the latitude and longitude coordinates of the target airport using the bilinear interpolation method.

[0049] ECMWF raw data is in the form of a regular grid. The target airport location is usually located inside a grid cell rather than exactly on a grid point. Therefore, bilinear interpolation is needed to calculate the feature values ​​of the target location using the values ​​of the four surrounding grid points. By performing bilinear interpolation on the four nearest grid points around the target point in both longitude and latitude directions, a good balance between computational accuracy and efficiency can be achieved, making it suitable for spatial downscaling of meteorological grid data. Using the example above, the target airport runway center point has longitude and latitude of (30.5°N, 115.0°E). The meteorological feature values ​​corresponding to this point need to be extracted from a 0.1°×0.1° resolution grid. After interpolation, the feature values ​​corresponding to each time period are extracted according to the reporting time and forecast lead time, forming ECMWF meteorological feature time-series data matched to the target airport's coordinates, providing basic meteorological feature input for subsequent feature engineering.

[0050] In step S122, the CAMS gridded pollution forecast data obtained in step S110 is interpolated to the latitude and longitude coordinates of the target airport using the same bilinear interpolation method as in step S121. The original CAMS data is also in a regular grid format, with a spatial resolution typically of 0.4°×0.4°. Using the example above, the target airport runway center point has latitude and longitude of (30.5°N, 115.0°E). Pollution element values ​​corresponding to this point need to be extracted from the 0.4°×0.4° resolution grid, including PM2.5 concentration, PM10 concentration, aerosol optical thickness, nitrogen dioxide concentration, ozone concentration, and sulfate aerosol concentration. After interpolation, the element values ​​corresponding to each time period are extracted according to the reporting start time and forecast lead time, forming time-series data of each CAMS pollution element matched to the target airport coordinates, providing atmospheric pollution feature input for subsequent feature engineering.

[0051] In step S123, the topographic features of the target airport are extracted using spatial analysis methods based on the DEM topographic elevation data obtained in step S110. The DEM data records surface elevation information in the form of regular grid points. First, the elevation value corresponding to the center point of the target airport runway is extracted from the DEM grid points as the airport's base altitude. Second, a certain range of DEM grid points (such as a 3×3 or 5×5 grid point window) is selected around the target airport location. The slope and topographic relief features are calculated using the elevation values ​​of each grid point within the window. The slope feature reflects the degree of inclination of the ground surface and can be calculated using the third-order inverse distance squared weighted difference method. The topographic relief feature reflects the degree of topographic variation and can be characterized by the standard deviation of the elevation values ​​within the window. Based on the above example, the runway center point elevation value is extracted from the 30-meter resolution SRTMDEM data, and the slope and terrain relief are calculated based on the surrounding 5×5 grid window to obtain DEM terrain elevation data containing three features: elevation, slope and terrain relief, which provides static terrain condition input for the terrain-weather adaptive modulation module.

[0052] In step S124, the airport measured visibility label data obtained in step S110 is precisely matched and aligned with the three types of data processed in steps S121 to S123 in terms of time dimension. The airport measured visibility label data comes from METAR reports or AWOS automatic station observations, and each observation time is marked with a timestamp. ECMWF time series data and CAMS time series data are both forecast data. Each data sample contains two time dimensions: the start time and the forecast lead time. During matching, alignment must be performed according to the same target time (i.e., the forecast target time) to ensure that the ECMWF forecast elements, CAMS pollution forecast elements, and DEM topographic features at that time form a one-to-one correspondence with the measured visibility label. Based on the above example, if the visibility recorded in the METAR report at a certain time is 500 meters, then the forecast values ​​of each ECMWF meteorological element, the forecast values ​​of each CAMS pollution element, and the topographic features corresponding to that time need to be obtained simultaneously to form a complete training sample. After matching and alignment, each sample contains meteorological features, pollution features, terrain features, and corresponding measured visibility labels at that moment, forming a multi-source fusion sample dataset, which enables the deep learning model to learn the complex mapping relationship between the forecast field and the observation field.

[0053] In step S125, data quality control is performed on the time-aligned multi-source fusion sample dataset obtained in step S124, removing samples containing missing values. Since ECMWF data, CAMS data, or measured visibility data may have missing values ​​at specific times due to data interface anomalies, storage deficiencies, or observation equipment malfunctions, directly using samples containing missing values ​​for model training would affect the model's learning performance and prediction accuracy.

[0054] The missing value detection method involves iterating through each time sample in the multi-source fusion sample dataset and checking whether the sample contains a missing value for any element. If any of the ECMWF meteorological elements, CAMS pollution elements, or measured visibility labels are missing, the entire time sample is deleted, retaining only samples with complete elements. Based on the above example, within the time range of January 2021 to December 2025, approximately N complete samples remain after missing value removal (depending on data quality), forming a cleaned and complete sample dataset to ensure the reliability and completeness of the data used for subsequent modeling.

[0055] In step S126, Z-score standardization is performed on all continuous numerical variables in the cleaned complete sample dataset obtained in step S125. Since different meteorological and pollution elements have different physical dimensions and numerical magnitudes—for example, temperature is on the order of tens, air pressure is on the order of hundreds to thousands, and PM2.5 concentration is on the order of tens to hundreds—directly inputting data of different magnitudes into a deep learning model would lead to unbalanced gradient updates, affecting the model's convergence speed and prediction performance. Z-score standardization converts each variable into a standard distribution with a mean of 0 and a standard deviation of 1. The formula is: for a variable x, its standardized value is x'=(x-μ) / σ, where μ is the mean of the variable and σ is the standard deviation of the variable.

[0056] During standardization, the mean and standard deviation of the training set samples are used to standardize the training, validation, and test sets respectively, ensuring the consistency of data distribution. After standardization, all variables are transformed to a unified dimension, resulting in preprocessed multi-source heterogeneous data, providing standardized data input for subsequent feature engineering and model training.

[0057] In conjunction with the first aspect, the preprocessed multi-source data in step S130 undergoes feature splicing and fusion, specifically including: S131, the meteorological elements in the preprocessed ECMWF numerical forecast data, the pollution elements in the CAMS air pollution forecast data, and the DEM topographic elevation data are stitched together to obtain the basic feature vector.

[0058] S132, based on meteorological elements in continuous time intervals in ECMWF numerical forecast data, calculates the trend value and statistics of each meteorological element within the preset sliding time window, and obtains the statistical characteristics of the time-series sliding window.

[0059] S133, based on the physical mechanism of atmospheric extinction, extracts the derived features of cross-combination from ECMWF numerical forecast data and CAMS air pollution forecast data, and obtains the combined cross-features based on the physical mechanism.

[0060] S134 concatenates the basic feature vector, time-series sliding window statistical features, and combined cross features to form a complete feature vector set.

[0061] In step S131, the three types of preprocessed data are horizontally concatenated along their feature dimensions to form a basic feature vector. Specifically, the ECMWF numerical forecast data includes at least meteorological elements such as 2-meter temperature, 2-meter dew point temperature, 10-meter wind speed U / V component, sea level pressure, relative humidity, cloud cover, total precipitation, and ECMWF visibility forecast values; the CAMS air pollution forecast data includes at least pollution elements such as PM2.5 concentration, PM10 concentration, aerosol optical thickness, nitrogen dioxide concentration, ozone concentration, and sulfate aerosol concentration; and the DEM topographic elevation data includes at least the elevation value, slope characteristics, and topographic relief characteristics of the target airport runway location. The sum of the feature dimensions of the above three types of data constitutes the total dimension of the basic feature vector. Referring to the example above, the basic feature vector dimension consists of 8 ECMWF elements, 6 CAMS elements, and 3 topographic-assisted features, totaling 17 dimensions. This basic feature vector serves as the foundation for subsequent feature construction and model input, encompassing raw information from three dimensions: meteorology, pollution, and topography, providing comprehensive input information for deep learning models.

[0062] In step S132, for meteorological elements in consecutive time intervals of ECMWF numerical weather prediction data, the trend values ​​and statistics of each meteorological element within a preset sliding time window are calculated to capture the temporal evolution trend of visibility-related meteorological elements. The preset sliding time window includes, but is not limited to, a past 3-hour window (t-3 to t), a past 6-hour window (t-6 to t), and a past 12-hour window (t-12 to t). The trend values ​​include the difference values ​​of each meteorological element within the sliding time window, such as temperature change trends (current temperature minus temperature 3 hours ago) and air pressure change trends (current air pressure minus air pressure 6 hours ago), used to characterize the short-term direction of change of meteorological elements. The statistics include the mean, maximum, and minimum values ​​of each meteorological element within the sliding time window, such as the maximum relative humidity in the past 6 hours and the mean wind speed in the past 12 hours, used to characterize the overall level and extreme conditions of meteorological elements in the recent period. Based on the above example, if the target time for a sample is t, then using t-3 to t, t-6 to t, and t-12 to t as windows respectively, the changes and statistical values ​​of key meteorological elements such as temperature, humidity, wind speed, and air pressure within these windows are calculated, forming a time-series sliding window statistical feature. By introducing the time-series sliding window statistical feature, the model can perceive the evolution trend of meteorological elements, which helps to determine the direction of visibility changes and provides time-series information support for the formation and dissipation of low visibility weather.

[0063] The atmospheric extinction coefficient is a direct physical quantity that determines visibility. According to the Koschmieder visibility-extinction relationship, visibility is inversely proportional to the atmospheric extinction coefficient. Aerosol particles reduce light radiation through scattering and absorption, and increased relative humidity leads to hygroscopic growth of aerosols, increasing their scattering cross-section and thus enhancing the extinction effect. This effect is characterized by the aerosol scattering hygroscopic growth factor f(RH). The hygroscopic growth effect is particularly significant when the relative humidity exceeds approximately 80%.

[0064] Based on the above physical mechanism, the combined cross features in this step include at least two of the following features: (1) Temperature-dew point difference, calculated by the formula Td=T2m-D2m, where T2m is the 2-meter temperature in the ECMWF numerical weather prediction data and D2m is the 2-meter dew point temperature in the ECMWF numerical weather prediction data. The temperature-dew point difference characterizes the saturation level of the near-surface atmosphere. When this value is small, it indicates that the atmosphere is close to saturation, which is conducive to the formation of fog.

[0065] (2) Near-surface wind speed scalar, the calculation formula is Ws=sqrt(U10) 2 +V10 2 The values ​​are: U10 and V10, where U10 and V10 are the 10-meter zonal and 10-meter meridional wind speeds in the ECMWF numerical weather prediction data, respectively. Wind speed directly affects the transport and diffusion of pollutants, and thus affects visibility.

[0066] (3) The wet extinction coefficient is used to characterize the extinction effect caused by the hygroscopic growth of aerosols under high humidity conditions. Its calculation formula is as follows: f_hum_ext=RH×AOD / (1-RH / 100) Where RH is the relative humidity (%) in the ECMWF numerical weather prediction data, and AOD is the aerosol optical thickness (550 nm) in the CAMS air pollution forecast data. This formula is based on atmospheric extinction theory, using (1-RH / 100) as a high-humidity amplification factor, and coupling AOD and RH multiplicatively. When RH is large (e.g., RH>80%), the denominator (1-RH / 100) approaches 0, and f_hum_ext increases sharply, simulating the strong extinction effect caused by aerosol hygroscopic growth under high humidity conditions, directly pointing to the haze / fog conversion process.

[0067] Based on the above examples, when ECMWF forecast data shows RH=85% and CAMS forecast data shows AOD=0.6 at a certain time, the surrogate value of the wet extinction coefficient f_hum_ext=0.85×0.6 / (1-0.85)=0.51 / 0.15=3.4. This value is significantly higher than that under low humidity conditions, indicating that the aerosol hygroscopic growth effect is strong under the current atmospheric conditions, resulting in a large extinction coefficient and low visibility. When the temperature-dew point difference Td=2℃, it indicates that the near-surface atmosphere is close to saturation, which is conducive to fog formation. By constructing the above-mentioned combined cross-features based on physical mechanisms, the model can better understand the strong extinction effect caused by aerosol hygroscopic growth under high humidity conditions, as well as the fog formation process under near-saturation conditions, thereby improving the prediction accuracy of low visibility weather.

[0068] In step S134, the basic feature vector obtained in step S131, the temporal sliding window statistical features obtained in step S132, and the combined cross-features based on physical mechanisms obtained in step S133 are concatenated along the feature dimension to form a complete feature vector set. This complete feature vector set covers static terrain features, original meteorological and pollution forecast features, temporal evolution features, and combined cross-features guided by physical mechanisms, realizing a complete mapping from multi-source heterogeneous data to a unified feature space. Based on the above example, the complete feature vector set consists of approximately tens of features (depending on the specific feature configuration), with each time sample corresponding to a complete feature vector, covering multiple dimensions affecting visibility, such as meteorological conditions, pollution levels, terrain constraints, and temporal evolution. This provides information-rich and physically interpretable feature input for the subsequent meteorological adaptive residual correction network. After being organized chronologically, this complete feature vector set is used as input to the meteorological adaptive residual correction network for model training or prediction.

[0069] In conjunction with the first aspect, step S140 includes: S141 organizes the feature vector set into a mixed time series sequence containing historical data segments and future forecast segments in chronological order.

[0070] S142, the mixed time series is input into the meteorological adaptive residual correction network, and local time series patterns at different time scales are extracted by multiple parallel one-dimensional dilated convolutions with different dilation rates. The outputs of each parallel dilated convolution are then concatenated to obtain multi-scale time series features.

[0071] S143, add the multi-scale temporal features and the physical sensing location code element by element to obtain the temporal features of the fused location information.

[0072] S144: The temporal features of the fused location information are input into the Transformer encoder. After processing by a multi-head self-attention mechanism and a feedforward network, the long-range dependencies between different time points in the mixed temporal sequence are captured to obtain the encoder feature sequence.

[0073] S145, the encoder feature sequence is input into the terrain-meteorological adaptive feature modulation module. Based on the global temporal context of the DEM terrain elevation data and the encoder feature sequence, modulation parameters are generated. The encoder feature sequence is dynamically modulated using the modulation parameters to obtain the feature sequence after terrain-meteorological adaptive modulation.

[0074] S146, the time steps corresponding to the future forecast segments in the modulated feature sequence are extracted and input into the decoding output layer for mapping to obtain the visibility residual prediction values ​​for each time period.

[0075] S147: Obtain the original visibility forecast values ​​for each forecast period in the ECMWF numerical forecast data. Add the visibility residual forecast values ​​for each period to the original visibility forecast values ​​for each period to obtain the corrected visibility forecast values ​​for each period. Use these as the visibility forecast results for the target airport in the future for multiple periods and output them.

[0076] Step S141 organizes the complete feature vector set constructed in step S130 into a hybrid time-series sequence in chronological order. This sequence is composed of historical data segments and future forecast segments. The historical data segments consist of feature vectors for each time period of the past 24 hours (a total of 8 time periods, with a time resolution of 3 hours), used to provide the recent evolution trend of weather and pollution conditions. The future forecast segments consist of feature vectors for each time period from 0 to 72 hours (a total of 24 time periods, with a time resolution of 3 hours), used to receive multi-source forecast data for each future forecast time period. The historical data segments and future forecast segments are concatenated along the time dimension to form a hybrid time-series sequence of length 32 (8+24), with each time step containing the complete feature vector for the corresponding time. Using the example above, if the current reporting time is 00:00 on October 1, 2024, then the historical data segment contains feature vectors from 8 time periods over the past 24 hours (i.e., from 00:00 the previous day to 00:00 on the current day), and the future forecast segment contains feature vectors from 24 time periods over the next 72 hours (i.e., from 00:00 on the current day to 00:00 on October 4), together forming a mixed time series sample. By organizing historical data and future forecasts into the same sequence, the network can automatically learn the relationship between the current data trend and the correction of future forecast errors using a self-attention mechanism.

[0077] Step S142 involves organizing the mixed time series sequence obtained in step S141 (combined with...) Figure 2The complete sequence (in the image) is first linearly projected onto the hidden dimension D (D=64 in this embodiment) through a fully connected layer, resulting in a sequence representation of dimension L×D, where L is the length of the mixed temporal sequence (L=32). This sequence representation is then input into a multi-scale dilated convolutional layer, which consists of three parallel one-dimensional dilated convolutions with dilation rates of 1, 2, and 4, respectively. Each convolutional kernel has a size of 3 and 32 output channels. The convolutional kernel with a dilation rate of 1 has a receptive field of 3 consecutive time steps, used to extract short-term fluctuation patterns between adjacent time steps, such as rapid changes in temperature and humidity at the hourly level. The convolutional kernel with a dilation rate of 2 expands its receptive field to 5 time steps, used to extract evolution patterns at the diurnal scale, such as the diurnal temperature and humidity cycle. The convolutional kernel with a dilation rate of 4 expands its receptive field to 9 time steps, used to extract persistent change patterns at the weather process scale, such as the passage of a cold front or persistent fog. The outputs of the three parallel convolutions are concatenated along the feature channel dimension to obtain a multi-scale temporal feature of dimension L×96, which is then output after layer normalization. By using multi-scale dilated convolution, the network can capture meteorological evolution patterns across different time spans simultaneously without increasing the number of parameters, effectively improving the efficiency of feature extraction and representation capabilities.

[0078] In step S143, the multi-scale temporal features obtained in step S142 are added element-wise to the physical perception location codes constructed in steps S1431 to S1434. The physical perception location codes have the same dimension as the multi-scale temporal features (L×96). By adding the multi-scale temporal features to the physical perception location codes, the network can utilize location information and physical prior knowledge in subsequent attention calculations to enhance its understanding of day-night and seasonal cycles.

[0079] In step S144, the temporal features of the fused location information obtained in step S143 are input into the Transformer encoder. In this embodiment, the Transformer encoder has two stacked layers, each containing a multi-head self-attention sublayer and a feedforward network sublayer. The multi-head self-attention mechanism contains four attention heads, the feedforward network has a hidden layer dimension of 256, and a dropout rate of 0.1. In the multi-head self-attention mechanism, each time step in the sequence can pay attention to the information of all other time steps in the sequence, thereby capturing the long-range dependencies between different time steps. Since the mixed time series sequence contains both historical real-time segments and future forecast segments, the self-attention mechanism can automatically learn the relationship between historical real-time trends and future forecast errors, as well as the interdependencies between different forecast lead times. For example, there is a strong correlation between the visibility residual at the current moment and the humidity change trend over the past 6 hours. The self-attention mechanism can automatically identify this cross-time-time dependency pattern without the need for manual rule design. After processing by the Transformer encoder, the encoder feature sequence H_enc is obtained, with a dimension of L×96.

[0080] In step S145, the encoder feature sequence obtained in step S144 is input into the terrain-meteorological adaptive feature modulation module. The input to this module consists of two parts: first, the DEM terrain elevation data obtained in step S123 (including elevation, slope, and terrain relief, with a dimension of 3, becoming 4-dimensional after adding the bias term); and second, the encoder feature sequence H_enc. This modulation mechanism allows the same terrain to play different modulating roles under different weather conditions: for example, when the global temporal context is characterized by high humidity and calm winds, the value of the humidity-sensitive channel in γ increases, enhancing the modulation effect of terrain on the humidity field and extinction effect; when the global temporal context is characterized by strong winds, this channel is suppressed, weakening the constraint effect of terrain on atmospheric flow. Through this modulation mechanism, the effect of terrain influence dynamically adapting to the weather background is achieved.

[0081] In step S146, feature vectors (i.e., the last 24 positions in the sequence) corresponding to 24 time steps are extracted from the terrain-meteorological adaptive modulation feature sequence H_mod obtained in step S145. Each time step corresponds to a forecast lead time. The extracted feature vectors for each forecast lead time are input into the decoding output layer for mapping. The decoding output layer is a two-layer fully connected network with a structure of 96→64→1, and the middle layer contains a ReLU activation function. Each forecast lead time shares the same decoding output layer weights. After mapping by the decoding output layer, the visibility residual prediction value Δ corresponding to each forecast lead time is output. _t, where t=1,2…24, corresponding to the residual prediction values ​​for forecast lead times of 3 hours, 6 hours…72 hours, respectively.

[0082] In step S147, the raw visibility forecast values ​​VIS_EC_t (t=1,2…24) for each forecast lead time are obtained from the ECMWF numerical weather prediction data. The visibility residual prediction values ​​Δ obtained in step S146 for each lead time are then compared. The original visibility forecast value VIS_EC_t for each time period is added to _t over time, and the corrected visibility forecast value for each time period is obtained according to the following formula, which is used as the visibility forecast result: VIS_corrected_t=VIS_EC_t+Δ _t, where t=1,2…24, corresponding to forecast lead times from 3 hours to 72 hours. Referring to the example above, if the original ECMWF visibility forecast for a certain lead time is 800 meters, and the residual value predicted by the network is -150 meters, then the corrected visibility forecast value, as the visibility prediction result, is 650 meters, which is closer to the actual visibility at the airport. Through the residual learning strategy, the network only needs to learn the local error portion that is difficult to resolve in the ECMWF forecast, reducing the learning difficulty and improving the generalization ability. Finally, the corrected visibility forecast values ​​for each lead time are output as the visibility forecast results for the target airport in the future for multiple lead times (0 to 72 hours, in 3-hour intervals), providing refined, multi-lead time visibility forecast services for aviation operations.

[0083] In conjunction with the first aspect, S143 includes: S1431: Obtain the absolute position of each time step in the mixed time series sequence, and generate a relative position encoding vector through sinusoidal position encoding.

[0084] S1432: Obtain the real-world time corresponding to each time step, calculate the solar hour angle based on the real-world time to represent day and night information, and calculate the annual day to represent seasonal information.

[0085] S1433 performs sine coding on the solar hour angle and the annual day, respectively, and maps the coding results to the hidden dimension through a fully connected layer to obtain the physical time coding vector.

[0086] S1434, add the relative position encoding vector to the physical time encoding vector element by element to obtain the physical sensing position encoding.

[0087] In this embodiment, the physical sensing location encoding is constructed through the following steps S1431 to S1434. Its dimension is the same as that of the multi-scale temporal feature, which is L×96 (L=32, D=96), so that the two can be directly added element by element.

[0088] In step S1431, the absolute position index pos (pos=0,1…31) of each time step in the mixed time series is obtained, and a relative position encoding vector is generated using the sinusoidal position encoding formula. The calculation formula for sinusoidal position encoding is: For position pos and dimension index i: PE(pos,2i)=sin(pos / 10000 2i / D ) PE(pos,2i+1)=cos(pos / 10000 2i / D ) Where D is the hidden dimension (D=96 in this embodiment), and i is the dimension index. Sine positional encoding utilizes sine and cosine functions of different frequencies to generate a unique encoding vector for each absolute position in the sequence, enabling the network to perceive the sequential order of each time step within the sequence. This encoding method is consistent with the positional encoding mechanism in the Transformer architecture, exhibiting good extrapolation capabilities and stable gradient propagation characteristics. Through absolute positional encoding, the network can distinguish between historical real-world segments (the beginning of the sequence) and future predicted segments (the end of the sequence), understanding the relative order between time steps.

[0089] In step S1432, the actual world time (UTC) corresponding to each time step in the mixed time series is obtained, that is, the real calendar time of the historical moment or the predicted future moment represented by that time step. For each time step, two physical time features are calculated based on its actual world time: The solar hour angle refers to the angle of the sun relative to due south as seen from an observation point, used to characterize day and night information. The solar hour angle is calculated as follows: first, the local solar time is calculated based on the UTC time and longitude corresponding to the time step; then, the solar hour angle is calculated using the local solar time. The solar hour angle ranges from -180° to 180°, where 0° represents noon, negative values ​​represent morning, positive values ​​represent afternoon, ±90° represents the vicinity of sunrise / sunset, and ±180° represents midnight. Using the solar hour angle, the network can perceive its current position within the day and night cycle, and understand the impact of light conditions on atmospheric stability and fog formation.

[0090] The cumulative day (CSB) refers to the ordinal number of days in a year for a given time step (from 1 to 365 or 366), used to represent seasonal information. The CSB is calculated by determining the ordinal number of that date within the current year based on the month and day information corresponding to that time step. Through the CSB, the network can perceive its seasonal position within the year and understand the impact of seasonal atmospheric circulation patterns, such as frequent winter inversions and active summer convection, on visibility.

[0091] In step S1433, the solar hour angle and annual day obtained in step S1432 are sine-encoded. Since both the solar hour angle and annual day are periodic variables (the solar hour angle has a period of 24 hours, and the annual day has a period of one year), directly using the original values ​​would make it difficult for the network to capture their periodic characteristics. Therefore, sine encoding is needed to map them to a cyclic space. The specific encoding method is as follows: the solar hour angle and annual day are used as angle values, and their sine and cosine values ​​are calculated to obtain the four-dimensional original physical time code (solar hour angle sine, solar hour angle cosine, annual day sine, and annual day cosine). Taking a certain time step of Ezhou Airport as an example, if the time step is 06:00 UTC on June 15, 2024 (corresponding to 14:00 local time), the solar hour angle is about 30°, and the yearly day is the 167th day, then the original physical time code is [sin30°,cos30°,sin(2π×167 / 365),cos(2π×167 / 365)]=[0.5,0.866,0.831,-0.557].

[0092] Subsequently, the original physical time encoding is input into a fully connected layer, which maps the four-dimensional original physical time encoding to a hidden dimension D=96, resulting in a 96-dimensional physical time encoding vector. The weights of this fully connected layer are optimized along with other network parameters during training, enabling the network to automatically learn how to incorporate physical time information into its feature representation.

[0093] In step S1434, the relative position encoding vector generated in step S1431 is added element-wise to the physical time encoding vector generated in step S1433 to obtain the final physical-aware position encoding. Both the relative position encoding vector and the physical time encoding vector have a dimension of 96, and the dimension remains 96 after addition. This physical-aware position encoding integrates absolute position information and prior physical time information: on the one hand, the relative position encoding provides the order information of time steps in the sequence, enabling the network to distinguish between the past and the future; on the other hand, the physical time encoding injects day-night and seasonal cycle information, enabling the self-attention mechanism to understand prior knowledge such as "high incidence of radiation fog in the early morning" and "frequent temperature inversions in winter," elevating the semantics of the time series from a simple sequence to a physically meaningful temporal context.

[0094] Finally, the multi-scale temporal features (dimension L×96) obtained in step S142 are added element-wise to the physical sensing location encoding (dimension L×96) to obtain the temporal features fused with location information. This feature, while preserving the multi-scale temporal pattern, incorporates sequential location information and diurnal / seasonal prior information, serving as input to the Transformer encoder. Referring to the example above, for the time step of 04:00 AM (solar hour angle approximately -60°), the solar hour angle component in the physical sensing location encoding has a relatively large value, enabling the network to perceive that this time is nighttime, which is beneficial for identifying the formation conditions of radiation fog. For the time step where the cumulonimbus day is in winter (day 10), the cumulonimbus component allows the network to perceive that it is currently in a season with frequent winter inversions, enhancing its sensitivity to winter fog. By introducing the physical sensing location encoding, the network can better utilize diurnal and seasonal prior knowledge, improving the prediction accuracy of low-visibility weather.

[0095] In conjunction with the first aspect, step S145, which involves dynamically modulating the encoder feature sequence using modulation parameters to obtain the terrain-weather adaptively modulated feature sequence, includes: S1451, input the DEM terrain elevation data into the terrain-meteorological adaptive feature modulation module to obtain the static terrain feature vector.

[0096] S1452, the encoder feature sequence is processed by global average pooling to obtain the global temporal context vector.

[0097] S1453 concatenates the static terrain feature vector with the global temporal context vector to obtain the fused feature vector.

[0098] S1454 inputs the fused feature vector into the fully connected network for mapping, and outputs a scaling parameter sequence and a shift parameter sequence.

[0099] S1455, apply the scaling parameter sequence and shift parameter sequence to each feature channel of the encoder feature sequence respectively, and perform channel-by-channel modulation on the encoder feature sequence to obtain the feature sequence after terrain-weather adaptive modulation.

[0100] In step S145, the encoder feature sequence obtained in step S144 is input into the terrain-weather adaptive feature modulation module, combined with... Figure 3 As shown, modulation parameters are generated using the global temporal context of DEM terrain elevation data and encoder feature sequences, and the encoder feature sequences are dynamically modulated to achieve adaptive adjustment of terrain influence with weather background.

[0101] In step S1451, the DEM terrain elevation data obtained in step S123 is input into the terrain-weather adaptive feature modulation module to obtain a static terrain feature vector. The DEM terrain elevation data includes at least the elevation value (H), slope feature (S), and terrain relief feature (T) of the target airport runway location, which together constitute a three-dimensional terrain feature vector. In this embodiment, a bias term (constant 1) is added to this three-dimensional feature vector to form a four-dimensional static terrain feature vector S=[H,S,T,1] T The dimension is 4. Based on the example above, the runway center point elevation is approximately 46 meters, the surrounding terrain undulation is approximately 5-10 meters, and the slope is approximately 0.5°-2°. Therefore, the static terrain feature vector S = [46, 1.2, 7.5, 1]. T The specific value depends on the actual terrain conditions. The purpose of adding a bias term is to enable the subsequent fully connected network to learn an adjustable reference offset when generating modulation parameters, thereby enhancing the network's expressive power.

[0102] In step S1452, the encoder feature sequence H_enc obtained in step S144 (with dimensions L×D, where L=32 is the length of the mixed temporal sequence and D=96 is the hidden dimension) is subjected to global average pooling along the time dimension to obtain the global temporal context vector C. The formula for calculating global average pooling is: C = (1 / L) × ΣH_enc[i,:] L Here, H_enc[i,:] represents the feature vector at the i-th time step in the encoder feature sequence. Through global average pooling, the encoder feature sequence is compressed into a fixed-length vector C with dimension D=96. This global temporal context vector comprehensively reflects the overall meteorological and pollution status characteristics of the entire mixed temporal sequence (including the past 24 hours and the next 72 hours), such as the overall average humidity level, average temperature level, and average pollution concentration level of the sequence. By using the global temporal context as a modulation condition, the network can dynamically adjust the modulation intensity of terrain features on the feature channels according to the current overall weather background, enabling the same terrain to play different regulatory roles under different weather conditions.

[0103] Step S1453 concatenates the static terrain feature vector S (dimension 4) obtained in step S1451 with the global temporal context vector C (dimension 96) obtained in step S1452 along the feature dimensions to obtain a fused feature vector F=[S;C] (dimension 100). This fused feature vector integrates static terrain conditions with dynamic weather background information and serves as the conditional input for subsequent modulation parameter generation. The concatenation operation enables sufficient feature interaction between terrain features and meteorological context features in the fully connected network, thereby generating modulation parameters adapted to the current weather background.

[0104] In step S1454, the fused feature vector F (dimension 100) obtained in step S1453 is input into a small fully connected network for mapping, outputting a scaling parameter sequence γ and a shift parameter sequence β. This fully connected network has a two-layer structure: the first layer maps the 100-dimensional input to 64 dimensions using the ReLU activation function; the second layer maps the 64-dimensional input to 96 dimensions, outputting the scaling parameter γ and the shift parameter β. Both γ and β have a dimension of 96, the same as the number of channels in the encoder feature sequence. Specifically, the mapping relationship of the first fully connected layer is: F1 = ReLU(W1·F + b1), where W1 is a 100×64 weight matrix and b1 is a 64-dimensional bias; the mapping relationship of the second fully connected layer is: γ,β = W2·F1 + b2, where W2 is a 64×192 weight matrix (output channel number is 96×2=192), and b2 is a 192-dimensional bias. Each element of γ controls the amplification or attenuation intensity of the corresponding feature channel (amplification when γ>1, attenuation when 0<γ<1), and each element of β controls the reference value offset of the corresponding feature channel.

[0105] Based on the above example, when the global time-series context exhibits meteorological conditions of high humidity, calm winds, and high pollution (e.g., RH=85%, wind speed <2m / s, PM2.5 concentration >75μg / m³), 3 The values ​​of the humidity-sensitive and pollution-sensitive channels in the output γ of the fully connected network are relatively large (e.g., γ_j>1.5), which enhances the representation strength of these channels and amplifies the constraint effect of terrain on the humidity field and pollutants in feature transmission. When the global temporal context shows strong wind and low humidity meteorological conditions, the values ​​of the corresponding channels in the output γ of the fully connected network are suppressed (e.g., γ_j<0.8), weakening the constraint effect of terrain on atmospheric flow, so that the modulated features better reflect the dominant role of large-scale weather background.

[0106] In step S1455, the scaling parameter sequence γ (96-dimensional) and shift parameter sequence β (96-dimensional) generated in step S1454 are applied to each feature channel of the encoder feature sequence H_enc, respectively, and the encoder feature sequence is modulated channel by channel according to the following formula: H_mod = γ⊙H_enc + β, where ⊙ denotes element-wise multiplication. For each time step t (t=1…L) and each feature channel c (c=1…D=96) in the encoder feature sequence, the modulated feature value is calculated as: H_mod[t,c] = γ_c × H_enc[t,c] + β_c. γ_c controls the scaling (enlargement or attenuation) of the c-th feature channel, and β_c controls the offset of the c-th feature channel. After channel-wise modulation, the terrain-weather adaptively modulated feature sequence H_mod is obtained, with the dimension still L×96.

[0107] This modulation mechanism allows the same terrain to play different regulatory roles under different weather conditions: when the meteorological context is high humidity and calm wind, the value of the humidity-sensitive channel in γ increases, and the constraint effect of terrain on local humidity distribution is amplified; when there is strong wind, the channel is suppressed, and the influence of terrain is weakened accordingly.

[0108] Based on the above example, the airport is located in the middle reaches of the Yangtze River Plain, with generally flat terrain but surrounded by waterways and a few hills. Under calm, stable, and humid weather conditions (such as autumn nights), topography has a significant impact on the formation of local fog. In this case, the humidity-sensitive channel in γ has a larger value, amplifying the modulating effect of topographic features on the humidity field. Under strong wind conditions (such as the passage of cold air in spring), the constraint effect of topography on atmospheric flow is weaker. In this case, the corresponding channel in γ has a smaller value, weakening the influence of topography. Through the above dynamic modulation mechanism, the network can adaptively adjust the contribution of topographic features to visibility prediction according to different weather backgrounds, thereby improving the accuracy and generalization ability of visibility prediction under complex meteorological conditions.

[0109] In conjunction with the first aspect, step S146 includes: S1461: Extract feature vectors corresponding to each forecast lead time of the future forecast segment from the feature sequence after terrain-meteorological adaptive modulation.

[0110] S1462 inputs the feature vector of each forecast time into the decoding output layer, and after passing through two fully connected networks for time-by-time mapping, outputs the visibility residual prediction value corresponding to each forecast time.

[0111] In step S146, the time steps corresponding to the future forecast segments are extracted from the feature sequence H_mod obtained in step S145 after terrain-meteorological adaptive modulation, and the visibility residual prediction values ​​for each time period are obtained by mapping through the decoding output layer. This is achieved through the following steps S1461 to S1462. Specifically: In step S1461, feature vectors corresponding to future forecast segments are extracted from the topographic-meteorological adaptively modulated feature sequence H_mod (dimension L×D, where L=32 is the total length of the hybrid time series sequence and D=96 is the hidden dimension). The hybrid time series sequence is composed of historical data segments and future forecast segments: the first 8 time steps in the sequence correspond to the historical data segments (past 24 hours, incremented by 3 hours), and the last 24 time steps correspond to the future forecast segments (0 to 72 hours, incremented by 3 hours).

[0112] Therefore, the feature vectors of the last 24 time steps in H_mod are taken and denoted as H_future=H_mod[8:32,:], with a dimension of 24×96, which correspond to 24 forecast lead times of 3 hours, 6 hours...72 hours respectively.

[0113] Based on the above example, if the current forecast start time is 00:00 on October 1, 2024, then the 24 extracted feature vectors correspond to the feature representations of 24 forecast times, from 03:00 and 06:00 on October 1, 2024 to 00:00 on October 4, 2024. Each feature vector integrates the meteorological conditions, pollution status, temporal evolution pattern, and channel enhancement information after terrain-meteorological adaptive modulation at that time.

[0114] In step S1462, the feature vectors of the 24 forecast timeframes extracted in step S1461 are input into the decoding output layer for mapping. The decoding output layer is a two-layer fully connected network: the first layer maps the 96-dimensional input features to 64 dimensions, using the ReLU activation function to introduce nonlinear transformation capability; the second layer further maps the 64-dimensional features to 1 dimension, outputting the visibility residual prediction value for that timeframe. Each forecast timeframe shares the same decoding output layer weights, meaning the same two-layer fully connected network processes the feature vectors of the 24 forecast timeframes sequentially, outputting the residual prediction value for each timeframe. The shared weight design keeps the network parameter size fixed, not increasing with the forecast timeframe, effectively reducing the risk of overfitting, and enabling the use of samples from all timeframes to jointly optimize the decoding layer parameters. After mapping by the decoding output layer, the visibility residual prediction value Δ for each timeframe is obtained. _t, where t=1,2…24, corresponds to the residual prediction values ​​for forecast lead times from 3 hours to 72 hours. Taking a certain forecast lead time for Ezhou Airport as an example, if the input to the decoding output layer is a feature vector (96-dimensional) for a certain forecast lead time, after two layers of mapping, a scalar value is output, which is the visibility residual prediction value for that lead time. This value may be positive (indicating that the ECMWF forecast is too low and needs to be corrected upwards) or negative (indicating that the ECMWF forecast is too high and needs to be corrected downwards). By extracting the feature vectors of future forecast segments and mapping them time-by-time through the decoding output layer, the network can independently generate residual prediction values ​​for each forecast lead time, achieving full-lead-time visibility residual prediction from 0 to 72 hours.

[0115] In step S150, before using the meteorological adaptive residual correction network for actual forecasting, network training needs to be completed in advance. Specifically: Based on the acquisition of ECMWF numerical forecast data, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data for the target airport within a historical period, a set of feature vectors for training is generated according to the same data preprocessing and feature engineering process as steps S120 and S130.

[0116] Obtain the visibility forecast value VIS_EC_t for each time period from the ECMWF numerical weather prediction data, obtain the measured visibility value VIS_OBS_t for the corresponding time period from the airport measured visibility label data, calculate the difference between the measured visibility value and the ECMWF visibility forecast value, and use this difference as the ground truth value for the residual training target of each time period, so that the network learns the error law of ECMWF forecast instead of relearning the complete physical process of visibility.

[0117] A meteorological adaptive residual correction network was constructed and its parameters were initialized. The network structure was configured as follows: the feature projection layer maps the input features to a hidden dimension D=64; the multi-scale dilated convolutional layer uses three parallel one-dimensional dilated convolutions with dilation rates [1,2,4], each with a kernel size of 3, and each convolution outputs 32 channels. The outputs of the three convolutions are concatenated along the channel dimension to obtain 96-dimensional features, which are then normalized; in the physical perception location encoding layer, the physical time encoding dimension is set to 16, and after being fused with the relative location encoding, it is projected to 96 dimensions through a fully connected layer; the Transformer encoder has two stacked layers, each containing four multi-head self-attention heads. The mechanism involves a feedforward network with a hidden layer dimension of 256 and a dropout rate of 0.1. The terrain-meteorological adaptive feature modulation module (FiLM modulator) receives a 100-dimensional fusion feature obtained by concatenating a 4-dimensional static terrain feature vector (including a bias term) with a 96-dimensional global temporal context vector. This fusion feature is mapped through a two-layer fully connected network (100→64→64) to output a 96-dimensional scaling parameter γ and a 96-dimensional shift parameter β. The decoding output layer adopts a two-layer fully connected network structure 96-64-1, with each forecast timeframe sharing the same network weights, and outputs the visibility residual prediction values ​​for each timeframe.

[0118] The training feature vector set is divided according to time sequence, with the samples from the first 80% of the time period used as the training set and the samples from the last 20% of the time period used as the test set; then, 20% of the samples from the first 80% of the training set are extracted in time sequence as the validation set.

[0119] The AdamW optimizer was used to optimize the network parameters, with an initial learning rate of 0.001 and a weight decay coefficient of 0.01 to prevent overfitting. The batch size was set to 16. The training feature vectors from the training set were batch-wise input into the meteorological adaptive residual correction network for forward propagation. After processing by each module of the network in steps S141 to S147, the network outputs the visibility residual prediction value Δ for each time period. The Huber loss function is used as the optimization objective. The loss function value is calculated based on the predicted visibility residual and the ground truth value of the residual training objective. The Huber loss combines the advantages of mean squared error and mean absolute error, exhibiting good robustness to outliers, and is particularly suitable for extreme low visibility samples that may exist in meteorological data. The gradients of the parameters in each layer are calculated through backpropagation based on the loss function value. The AdamW optimizer is then used to update the network parameters, completing one round of iterative training.

[0120] During training, a cosine annealing learning rate scheduling strategy is employed, causing the learning rate to periodically decay from an initial value of 0.001 to near zero as the number of training epochs increases. This helps the model more precisely approximate the optimal solution in the later stages of training. The process of forward propagation, loss calculation, and backpropagation to update parameters is repeated until the preset maximum number of training epochs (200 epochs) is reached. After each training epoch, the validation set is input into the current network to calculate the validation loss. Training is terminated early if the validation loss does not decrease within 20 consecutive training epochs to prevent overfitting. After training is complete, the network parameters with the smallest loss value on the validation set are saved as the final model parameters.

[0121] Through the above training process, the network learns the complex mapping relationship from multi-source features to visibility residuals, achieving high-precision correction of ECMWF visibility forecasts.

[0122] After completing model training, multi-time visibility prediction and model evaluation are performed on the test set samples to verify the prediction accuracy and effectiveness of the proposed method.

[0123] (1) Prediction output: Load the model parameters with the minimum loss in the validation set saved in step S150 to construct the trained meteorological adaptive residual correction network. Input the feature vectors in the test set into the loaded network, and perform forward propagation according to the process of steps S141 to S147 to output the visibility residual prediction values ​​of the target airport from 0 to 72 hours (3 hours each, for a total of 24 time periods). Add the visibility residual prediction values ​​of each time period to the original visibility forecast values ​​of ECMWF for each time period to obtain the corrected visibility prediction values ​​for each time period, which are used as the final multi-time period visibility prediction results.

[0124] (2) Calculation of evaluation indicators: Root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) were calculated for each forecast lead time (3 hours, 6 hours...72 hours). Among them, RMSE is used to measure the overall deviation between the predicted value and the observed value, and is more sensitive to larger errors; MAE is used to measure the average magnitude of the prediction error, reflecting the intuitive level of prediction accuracy; correlation coefficient R is used to measure the linear correlation between the predicted value and the observed value, reflecting the model's ability to capture the trend of visibility changes.

[0125] The prediction results of our proposed method (Meteorological Adaptive Residual Correction Network) were compared and evaluated with the following benchmark methods: raw ECMWF visibility forecasts (raw outputs without any corrections), XGBoost models, LightGBM models, random forest models, and linear regression models. All benchmark methods were trained and evaluated using the same training and test sets to ensure fairness in the comparison. By comparing RMSE, MAE, and R indices at different time points, the accuracy advantage of our proposed method in airport single-point multi-time-leader visibility forecasting tasks was verified, particularly its forecast improvement effect under low visibility conditions.

[0126] Secondly, embodiments of this application provide an airport visibility prediction device based on machine learning, combined with... Figure 4 As shown, the device includes: an acquisition module 10, a preprocessing module 20, a fusion module 30, and a prediction module 40.

[0127] The acquisition module is used to acquire multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data.

[0128] The preprocessing module is used to preprocess multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing value processing and standardization preprocessing.

[0129] The fusion module is used to perform feature splicing and fusion on the preprocessed multi-source data to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms.

[0130] The prediction module is used to input the feature vector set into a pre-trained meteorological adaptive residual correction network and output the visibility prediction results of the target airport in the future multiple time periods.

[0131] The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected sequentially. Based on the ECMWF visibility forecast, the network extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and then uses the terrain features extracted from the DEM terrain elevation data and the global context vector output by the Transformer encoder to generate modulation parameters, dynamically modulating the encoder output sequence. The decoding output layer then decodes and outputs the visibility residual prediction values ​​for each time period. These residual prediction values ​​are added to the original ECMWF visibility forecast values ​​to obtain the corrected visibility prediction values, which serve as the visibility prediction results.

[0132] Thirdly, embodiments of this application provide an electronic device, combined with Figure 5 As shown, the electronic device includes a memory 131 and a processor 130. The memory 131 stores a computer program, and the processor 130 runs the computer program to make the electronic device perform the above-described method.

[0133] Furthermore, combined Figure 5 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0134] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0135] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0136] Fourthly, embodiments of this application provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0137] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0138] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0139] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0141] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based airport visibility prediction method, characterized in that, The method includes: Acquire multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data; The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing measurement value processing and standardization preprocessing. The preprocessed multi-source data is spliced ​​and fused to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms. The feature vector set is input into a pre-trained meteorological adaptive residual correction network, which outputs the visibility prediction results of the target airport in the future multiple time periods. The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected in sequence. The network uses the ECMWF visibility forecast as a baseline. It extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and uses terrain features extracted from the DEM terrain elevation data and the global context vector output by the Transformer encoder to generate modulation parameters. This dynamically modulates the encoder output sequence. The decoding output layer decodes and outputs the visibility residual prediction values ​​for each time period. These residual prediction values ​​are then added to the original ECMWF visibility forecast value to obtain the corrected visibility prediction value, which serves as the visibility prediction result.

2. The method according to claim 1, characterized in that, The step of preprocessing the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data includes: The ECMWF numerical forecast data is interpolated to the latitude and longitude coordinates of the target airport using a bilinear interpolation method to obtain time series data of each meteorological element of ECMWF matched to the target airport; The CAMS air pollution forecast data is interpolated to the latitude and longitude coordinates of the target airport using a bilinear interpolation method to obtain time series data of each pollution element of CAMS matched to the target airport; The elevation values ​​of the target airport runway location are extracted from the DEM topographic elevation data, and the slope characteristics and topographic relief characteristics are calculated based on the DEM grid points of the area surrounding the target airport to obtain the DEM topographic elevation data of the target airport. The measured visibility label data of the airport is time-stamped and precisely matched and aligned with the interpolated ECMWF time series data, the CAMS time series data and the DEM terrain elevation data to obtain a time-aligned multi-source fusion sample dataset. Missing values ​​are detected in the multi-source fusion sample dataset, and time samples with missing values ​​are deleted to obtain a cleaned complete sample dataset. Z-score standardization is performed on all continuous numerical variables in the complete sample dataset to convert each variable to a uniform scale, resulting in preprocessed multi-source heterogeneous data.

3. The method according to claim 1, characterized in that, The steps for feature stitching and fusion of preprocessed multi-source data include: The meteorological elements in the preprocessed ECMWF numerical forecast data, the pollution elements in the CAMS air pollution forecast data, and the DEM topographic elevation data are stitched together to obtain the basic feature vector. Based on the meteorological elements of consecutive time periods in the ECMWF numerical forecast data, the changing trend value and statistics of each meteorological element within the preset sliding time window are calculated to obtain the time-series sliding window statistical characteristics. Based on the physical mechanism of atmospheric extinction, derived features of cross-combination are extracted from the ECMWF numerical forecast data and the CAMS air pollution forecast data to obtain combined cross features based on physical mechanisms; The basic feature vector, the time-series sliding window statistical features, and the combined cross features are concatenated to form a complete feature vector set.

4. The method according to claim 1, characterized in that, The steps of inputting the feature vector set into a pre-trained meteorological adaptive residual correction network and outputting visibility prediction results for the target airport in multiple future timeframes include: The feature vector set is organized in chronological order into a mixed time series sequence containing historical data segments and future forecast segments; The hybrid time series is input into the meteorological adaptive residual correction network, and local time series patterns at different time scales are extracted by multiple parallel one-dimensional dilated convolutions with different dilation rates. The outputs of each parallel dilated convolution are then concatenated to obtain multi-scale time series features. The multi-scale temporal features are added element by element to the physical sensing location code to obtain the temporal features fused with location information; The temporal features of the fused location information are input into the Transformer encoder. After processing by a multi-head self-attention mechanism and a feedforward network, the long-range dependencies between different time points in the hybrid temporal sequence are captured to obtain the encoder feature sequence. The encoder feature sequence is input into the terrain-meteorological adaptive feature modulation module. Modulation parameters are generated based on the global temporal context of the DEM terrain elevation data and the encoder feature sequence. The encoder feature sequence is dynamically modulated using the modulation parameters to obtain the terrain-meteorological adaptively modulated feature sequence. The time steps corresponding to the future forecast segments in the modulated feature sequence are extracted and input into the decoding output layer for mapping to obtain the visibility residual prediction values ​​for each time period. Obtain the original visibility forecast values ​​for each forecast period from the ECMWF numerical forecast data, add the visibility residual prediction values ​​for each period to the original visibility forecast values ​​for each period, and obtain the corrected visibility prediction values ​​for each period. These corrected visibility prediction values ​​are then used as the visibility prediction results for the target airport in the future for multiple periods and output.

5. The method according to claim 4, characterized in that, Before the step of adding the multi-scale temporal features to the physically sensed location code element by element to obtain the temporal features fused with location information, the method further includes: Obtain the absolute position of each time step in the hybrid time series sequence, and generate a relative position encoding vector through sinusoidal position encoding; Obtain the real-world time corresponding to each time step, calculate the solar hour angle based on the real-world time to represent day and night information, and calculate the annual day to represent seasonal information; The solar hour angle and the accumulated day of the year are sinusoidally encoded respectively, and the encoding results are mapped to the hidden dimension through a fully connected layer to obtain the physical time encoding vector. The relative position encoding vector is added element by element to the physical time encoding vector to obtain the physical sensing position encoding.

6. The method according to claim 4, characterized in that, The step of dynamically modulating the encoder feature sequence using the modulation parameters to obtain the terrain-weather adaptively modulated feature sequence includes: The DEM topographic elevation data is input into the topographic-meteorological adaptive feature modulation module to obtain a static topographic feature vector; The encoder feature sequence is subjected to global average pooling to obtain a global temporal context vector; The static terrain feature vector is concatenated with the global temporal context vector to obtain a fused feature vector; The fused feature vector is input into a fully connected network for mapping, and the output is a scaling parameter sequence and a shift parameter sequence; The scaling parameter sequence and the shift parameter sequence are applied to each feature channel of the encoder feature sequence respectively, and the encoder feature sequence is modulated channel by channel to obtain the feature sequence after terrain-weather adaptive modulation.

7. The method according to claim 4, characterized in that, The step of extracting the time step corresponding to the future forecast segment from the modulated feature sequence and inputting it into the decoding output layer for mapping to obtain the visibility residual prediction value for each time period includes: From the feature sequence after terrain-meteorological adaptive modulation, extract the feature vectors corresponding to each forecast lead time of the future forecast segment; The feature vectors of each forecast time are extracted and input into the decoding output layer. After passing through two fully connected networks, the visibility residual prediction value corresponding to each forecast time is output.

8. An airport visibility prediction device based on machine learning, characterized in that, The device includes: The acquisition module is used to acquire multi-source heterogeneous data, which includes at least: ECMWF numerical forecast data of the target airport, DEM topographic elevation data, CAMS air pollution forecast data, and airport measured visibility label data. The preprocessing module is used to preprocess the multi-source heterogeneous data to obtain preprocessed multi-source heterogeneous data; the preprocessing includes at least: spatiotemporal alignment, missing value processing and standardization preprocessing; The fusion module is used to perform feature splicing and fusion on the preprocessed multi-source data to construct a feature vector set that includes basic features, time-series sliding window statistical features, and combined cross features based on physical mechanisms. The prediction module is used to input the feature vector set into a pre-trained meteorological adaptive residual correction network and output the visibility prediction results of the target airport in the future multiple time periods. The meteorological adaptive residual correction network comprises a multi-scale dilated convolutional layer, a physical sensing location encoding layer, a Transformer encoder, a terrain-meteorological adaptive feature modulation module, and a decoding output layer, connected in sequence. Based on the ECMWF visibility forecast, the network extracts temporal local features through the multi-scale dilated convolutional layer, superimposes the location encoding generated by the physical sensing location encoding layer, processes it through the Transformer encoder, and uses terrain features extracted from the DEM terrain elevation data and the global context vector output by the Transformer encoder to generate modulation parameters, dynamically modulating the encoder output sequence. The decoding output layer then decodes and outputs the visibility residual prediction values ​​for each time period. These residual prediction values ​​are added to the original ECMWF visibility forecast values ​​to obtain the corrected visibility prediction values, which serve as the visibility prediction results.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to cause the electronic device to perform the method of any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores computer program instructions, which, when read and executed by a processor, perform the method described in any one of claims 1 to 7.

Images