Route ice accretion prediction method and device based on dynamic integration of multi-source data and model

Abstract

The application provides a route ice accretion prediction method and device based on multi-source data and model dynamic integration. The method comprises the following steps: acquiring multi-source data to construct an ice accretion sample data set, training and testing a plurality of base models through the ice accretion sample data set, obtaining the prediction accuracy of each base model for different ice accretion levels based on a confusion matrix, selecting, for each ice accretion level, the base model with the highest prediction accuracy as the expert model of the ice accretion level to obtain an expert model lookup table, acquiring prediction field data, constructing a pattern feature data set based on a target flight route, inputting the pattern feature data set into the plurality of trained base models to obtain a preliminary ice accretion level prediction result, and obtaining the final ice accretion level prediction result of the target flight route based on the preliminary ice accretion level prediction result and the expert model lookup table. The method effectively overcomes the shortcomings of generalization and roughness of traditional prediction results, and significantly improves the identification reliability and overall prediction accuracy of moderate and severe dangerous ice accretion.

Description

Technical Field

[0001] This application relates to the field of icing route forecasting technology, and in particular to a method, device, electronic device and storage medium for route icing forecasting based on dynamic integration of multi-source data and models. Background Technology

[0002] Aircraft icing is the phenomenon where water droplets on the aircraft's surface (such as wings and engine air intakes) freeze into ice as the aircraft passes through environments containing supercooled water droplets, such as clouds. Icing significantly increases aircraft weight, alters aerodynamic shape, impairs flight stability and maneuverability, and may cause navigation and communication equipment malfunctions, seriously threatening flight safety. With the continuous growth of civil aviation traffic and the increasing scarcity of airspace resources, unprecedented demands are placed on the accuracy and timeliness of en-route icing forecasts. Furthermore, the rapid development of the low-altitude economy presents new challenges to low-altitude flight safety. The variable weather and limited flight space in low-altitude environments exacerbate the hazards of icing, while traditional meteorological services geared towards high-altitude routes are insufficient to meet the refined needs of low-altitude flight.

[0003] Currently, numerical weather prediction models combined with traditional icing diagnostic algorithms (such as IC, CIP, and FIP) are the mainstream techniques for forecasting icing potential. However, these methods have significant limitations: First, traditional algorithms rely heavily on a limited number of meteorological factors (such as temperature and relative humidity), failing to fully utilize the increasingly rich cloud microphysical parameters (such as particle scale and vertical velocity) in numerical models. This results in low forecast accuracy, coarse spatial resolution, and weak ability to capture high-risk extreme icing events. Second, the algorithm outputs are not refined enough. For example, the FIP algorithm can only provide the probability of icing but cannot predict its intensity. In practical applications, this may lead to misjudgments of high probability but weak intensity or low probability but strong intensity, which is detrimental to accurate risk assessment and decision-making. Third, existing forecast products are mostly large-scale gridded weather maps, lacking refined, minute-level resolution targeted forecast products for fixed flight paths, making it difficult to provide pilots with intuitive and actionable avoidance advice.

[0004] On the other hand, icing is closely related to artificial rain enhancement operations. The potential areas for icing often overlap with cloud areas rich in supercooled water and with rain enhancement potential, which leads to a contradictory situation where rain enhancement aircraft have clouds to broadcast but cannot enter due to the risk of icing.

[0005] Therefore, developing forecasting technologies that can accurately identify icing risks and serve specific flight routes is of urgent practical significance for balancing the efficiency of artificial rain enhancement operations with flight safety, ensuring the safe conduct of low-altitude economic activities, and comprehensively improving the level of aviation meteorological services. Summary of the Invention

[0006] This application aims to at least partially address one of the technical problems in the related art.

[0007] Therefore, the first objective of this application is to propose a method for forecasting airway icing based on dynamic integration of multi-source data and models, so as to improve the accuracy of icing forecasts.

[0008] The second objective of this application is to propose a route icing forecasting device based on the dynamic integration of multi-source data and models.

[0009] The third objective of this application is to propose an electronic device.

[0010] The fourth objective of this application is to provide a computer-readable storage medium.

[0011] The fifth objective of this application is to provide a computer program product.

[0012] To achieve the above objectives, the first aspect of this application proposes a method for forecasting airway icing based on dynamic integration of multi-source data and models, comprising: A dataset of ice accumulation samples was constructed by acquiring multi-source data, including airborne detection time-series data, reanalysis data, and satellite remote sensing data. Multiple base models were trained and tested using the aforementioned ice accumulation sample dataset, and the prediction accuracy of each base model for different ice accumulation levels was obtained based on the confusion matrix. For each ice accumulation level, the base model with the highest prediction accuracy is selected as the expert model for that ice accumulation level, and an expert model lookup table is obtained. Acquire forecast field data, construct a pattern feature dataset based on the target flight path, input the pattern feature dataset into multiple trained base models, and obtain preliminary icing level prediction results; Based on the preliminary icing level prediction results, the final icing level prediction results of the target flight path are obtained by using an expert model lookup table.

[0013] In some implementations, the airborne detection time series data includes icing intensity, temperature, relative humidity, dew point temperature, liquid water content, cloud droplet number concentration, effective cloud droplet diameter, and atmospheric precipitable water; the reanalysis data includes vertical velocity, inversion layer intensity, and water vapor flux; and the satellite remote sensing data includes cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness.

[0014] In some implementations, the step of acquiring multi-source data to construct an ice accumulation sample dataset includes: The airborne detection time series data is acquired, and the airborne detection time series data is aggregated based on a preset time window and sliding step size to obtain the sample time, spatial location and feature vector of each time window; The ice accumulation intensity is divided into four ice accumulation intensity levels, and the ice accumulation intensity level with the highest frequency in each time window is used as the level label for that time window. Based on the alignment of the sample time and spatial location, the corresponding reanalysis data and satellite remote sensing data are extracted to obtain derived features and satellite regional features, respectively. The time, spatial location, feature vector, derived features, satellite region features, and grade labels of the samples are fused to form a single sample, and all the single samples constitute the ice accumulation sample dataset.

[0015] In some implementations, the step of training and testing multiple base models using the icing sample dataset, and obtaining the prediction accuracy of each base model for different icing levels based on the confusion matrix, includes: The ice accumulation sample dataset is divided into a training set and a test set; Each base model is trained using the training set. The prediction level labels of each base model are obtained using the test set; Based on the level labels and predicted level labels of each base model, a confusion matrix is ​​constructed to obtain the prediction accuracy of each base model for different ice intensity levels.

[0016] In some implementations, the base model includes random forest, XGBoost, and LightGBM.

[0017] In some implementations, the acquisition of forecast field data and the construction of a pattern feature dataset based on the target flight path include: Acquire operational numerical model forecast field data with spatial resolution at the hundred-meter level and temporal resolution at the minute level; Meteorological elements of the same type as the ice accumulation sample dataset are extracted from the forecast field data to obtain a three-dimensional meteorological element field. Determine the target flight route, which consists of multiple waypoints arranged in a time sequence. The waypoints include time, longitude, latitude, and altitude information. Meteorological feature vectors for each waypoint are extracted based on the three-dimensional meteorological element field. The feature vectors of all waypoints are arranged in order according to the target flight route to obtain the pattern feature dataset.

[0018] In some implementations, the extraction of meteorological feature vectors for each waypoint based on the three-dimensional meteorological element field includes: Based on the timestamps, longitudes, latitudes, and altitudes of the waypoints located in the three-dimensional meteorological field; Based on the longitude and latitude of the waypoints, spatial interpolation is performed in the three-dimensional meteorological element field to obtain the corresponding terrain height; Calculate the relative ground height based on the altitude and terrain elevation; Based on the relative ground height, the waypoint pressure value is estimated using the static equation and normalized to the Sigma layer coordinate value of the terrain-following coordinate system to obtain the waypoint time series including timestamp, longitude, latitude and Sigma layer coordinate value; Select the time data corresponding to the timestamp, obtain the horizontal meteorological element values ​​based on the longitude and latitude using bilinear interpolation, obtain the vertical meteorological element values ​​based on the Sigma layer coordinate values ​​using linear interpolation, and generate the meteorological feature vector based on the time data, horizontal meteorological element values ​​and vertical meteorological element values.

[0019] In some implementations, the preliminary icing level prediction result includes: For each waypoint, multiple icing level predictions are obtained based on multiple trained base models.

[0020] In some implementations, after obtaining the final icing level prediction result of the target flight path using an expert model lookup table based on the preliminary icing level prediction result, the method further includes: Based on the final icing level prediction results of all waypoints of the target flight route, a visualized icing potential forecast product is generated; the visualized icing potential forecast product includes: using geographic information as a base map, drawing the target flight route as a continuous curve, and marking the corresponding waypoint locations with different colors or graphic markers according to the final icing level prediction results of each waypoint.

[0021] To achieve the above objectives, a second aspect of this application proposes a route icing forecasting device based on dynamic integration of multi-source data and models, comprising: The data acquisition module is used to acquire multi-source data to construct an ice accumulation sample dataset, wherein the multi-source data includes airborne detection time series data, reanalysis data, and satellite remote sensing data; The model training module is used to train and test multiple base models using the ice accumulation sample dataset, and to obtain the prediction accuracy of each base model for different ice accumulation levels based on the confusion matrix. The model selection module is used to select the base model with the highest prediction accuracy for each ice accumulation level as the expert model for that ice accumulation level, and obtain the expert model lookup table. The model application module acquires forecast field data, constructs a pattern feature dataset based on the target flight path, and inputs the pattern feature dataset into multiple trained base models to obtain preliminary icing level prediction results. The icing prediction module uses an expert model lookup table based on the preliminary icing level prediction results to obtain the final icing level prediction results for the target flight path.

[0022] To achieve the above objectives, a third aspect of this application provides an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method described in the first aspect.

[0023] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, are used to implement the method described in the first aspect.

[0024] To achieve the above objectives, a fifth aspect of this application provides a computer program product including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0025] This application provides a method, device, electronic equipment, and storage medium for flight path icing forecasting based on dynamic integration of multi-source data and models. By deeply fusing multi-source data, it constructs an icing sample dataset encompassing macroscopic atmospheric background, cloud structure, and microscopic cloud physical states, providing a rich data foundation for model training. Based on a confusion matrix, it dynamically integrates and optimizes multiple models, and through in-depth analysis of the performance characteristics of different machine learning models in forecasting different levels of icing, it constructs a dynamic and reliable integrated forecasting system, significantly improving the overall accuracy of icing forecast levels. Employing a high-resolution numerical model, it performs precise spatiotemporal data mining and interpolation on the target flight path, transforming the gridded meteorological forecast field into a sequential feature dataset that is isomorphic to historical training data along a predetermined trajectory. This enables the direct conversion of grid point forecasts into customized, high-resolution numerical icing risk forecasts for specific flight paths. Finally, it transforms the digitized icing forecast results into intuitive, easy-to-understand, and directly serviceable graphical business products for flight decision-making, providing direct evidence for real-time flight decisions and achieving seamless integration between forecast products and user decision-making scenarios, greatly enhancing the practicality and operational efficiency of forecasts. This method effectively overcomes the shortcomings of traditional forecast results being generalized and crude, significantly improving the reliability of identifying moderate and severe hazardous icing and the overall forecast accuracy. It provides precise, intuitive, and operable key technical support for pilots to avoid risks, ensure low-altitude flight safety, and safely schedule aircraft for weather modification operations, and has high operational application value and socio-economic benefits.

[0026] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for forecasting airway icing based on dynamic integration of multi-source data and models, provided in an embodiment of this application; Figure 2 A two-dimensional schematic diagram of the visualized ice accumulation potential forecast product provided as an example in this application; Figure 3 A 3D schematic diagram of the visualized ice accumulation potential forecast product provided as an example in this application; Figure 4 A block diagram of a route icing forecasting device based on dynamic integration of multi-source data and models provided in this application embodiment; Figure 5 This is a block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0029] The following description, with reference to the accompanying drawings, describes a method, apparatus, and device for forecasting airway icing based on dynamic integration of multi-source data and models, according to embodiments of this application.

[0030] Figure 1 This is a flowchart illustrating a method for forecasting airway icing based on dynamic integration of multi-source data and models, provided in an embodiment of this application.

[0031] It should be noted that the execution subject of the route icing forecasting method based on multi-source data and dynamic model integration in this application embodiment is the route icing forecasting device based on multi-source data and dynamic model integration in this application embodiment. The route icing forecasting device based on multi-source data and dynamic model integration can be configured in an electronic device so that the electronic device can perform the route icing forecasting function based on multi-source data and dynamic model integration.

[0032] like Figure 1 As shown, this method for forecasting airway icing based on dynamic integration of multi-source data and models includes the following steps: Step S101: Obtain multi-source data to construct an ice accumulation sample dataset. The multi-source data includes airborne detection time series data, reanalysis data, and satellite remote sensing data.

[0033] As one implementation method, as shown in Table 1, the airborne detection time series data includes icing intensity, temperature, relative humidity, dew point temperature, liquid water content, cloud droplet number concentration, effective cloud droplet diameter, and atmospheric precipitable water; the reanalysis data includes vertical velocity, inversion layer intensity, and water vapor flux; and the satellite remote sensing data includes cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness.

[0034] Table 1. Data sources and corresponding meteorological elements

[0035] Airborne observation time-series data, serving as the core source of truth, comes from high-frequency (1 second / time) observations recorded by aircraft equipped with specialized meteorological observation equipment, such as aircraft used for artificial rain enhancement operations, during their flight along flight paths. The directly observed data includes: labeling information on whether icing has occurred and its intensity (none, light, moderate, severe), as well as the instantaneous microphysical and environmental conditions within the cloud during icing, such as temperature, relative humidity, liquid water content, cloud droplet number concentration, and effective cloud droplet diameter.

[0036] For the reanalysis data, the Global Reanalysis Dataset (ERA5) was used, providing atmospheric dynamic and thermal background field data covering the study area with a standard grid and fixed time intervals (1 hour / time). It not only provides basic fields such as temperature, humidity, and vertical velocity, but more importantly, it is used to calculate derived characteristics that indicate atmospheric stability and water vapor transport, such as inversion layer strength and overall water vapor flux.

[0037] For satellite remote sensing data, quantitative products (L2 level) from geostationary meteorological satellites (FY-4A / B) are used to provide high spatiotemporal resolution (15 min / time) information on macroscopic physical properties of clouds, such as cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness, to characterize the overall development and vertical structure of icing cloud systems.

[0038] As one implementation method, an ice accumulation sample dataset is constructed by acquiring data from multiple sources, including: Airborne detection time-series data is acquired. Based on a preset time window and sliding step size, the airborne detection time-series data is aggregated to obtain the sample time, spatial location, and feature vector for each time window. The icing intensity is divided into four icing intensity levels, and the icing intensity level with the highest frequency within each time window is used as the level label for that time window. Based on the alignment of sample time and spatial location, corresponding reanalysis data and satellite remote sensing data are extracted to obtain derived features and satellite regional features, respectively. The sample time, spatial location, feature vector, derived features, satellite regional features, and level label are fused to form a single sample. All single samples constitute the icing sample dataset.

[0039] In some embodiments, raw airborne sounding time-series data are collected and compiled, including records containing icing intensity labels and key microphysical meteorological elements. A fixed time window (5 min) and sliding step size (5 min) are set. The representative time of the window is (window start time + 2.5 min), and the representative spatial location of the window is the average of all latitude, longitude, and altitude within the window. The data is high-frequency and continuous, without external interpolation. The regional statistical strategy is to perform full sample statistics within the window, and the continuous airborne sounding time-series data are segmented. The window aggregates and statistically analyzes the average (typical state) and maximum (capturing peak risk, such as high liquid water leading to severe icing) values ​​of temperature, relative humidity, dew point temperature, liquid water content, cloud droplet number concentration, and effective cloud droplet diameter. The icing intensity label is taken as the mode; if there are multiple modes, the most severe level is taken.

[0040] Based on the obtained window representing time and spatial location, the most spatiotemporally matching data points are found in the ERA5 reanalysis data, with a time matching error of <30 minutes. Spatial matching prioritizes the nearest grid point; if the deviation is large, trilinear (latitude, longitude, altitude, and pressure layer) interpolation is performed. The regional statistical strategy does not employ regional expansion, but instead uses a point matching strategy with only single grid points or interpolation points. This is because the ERA5 reanalysis data represents the atmospheric background field, and its spatiotemporal resolution is not as fine as airborne data, but this does not affect the analytical application of the atmospheric background field. The required reanalysis elements are extracted, and their temperature, pressure, wind, and humidity fields are used to calculate derived characteristics such as inversion layer strength (reflecting atmospheric stratification stability) and overall water vapor flux (reflecting water vapor transport intensity) according to preset physical formulas, enriching the synoptic-scale information of the sample.

[0041] Based on the obtained window representing time and spatial location, the FY-4A / B satellite remote sensing data with the closest time is matched, with a time matching error of <7.5 min. Spatially, a 3×3 pixel region is extracted centered on the latitude and longitude of the location represented by the window, without interpolation. The regional statistical strategy is to calculate statistics for 9 pixels. Because satellite resolution is limited, point matching is easily affected by cloud edges. The average value (to smooth noise) and standard deviation (to quantify regional heterogeneity) of cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness within this region are calculated to characterize the macroscopic cloud characteristics of the sample point's region, while smoothing local noise in the satellite data.

[0042] All information corresponding to the same time window is integrated. This information mainly comes from aggregated airborne features, icing labels, derived features, and satellite regional features. The three types of data have different temporal resolutions: airborne detection time series data is 1 second, ERA5 reanalysis data is 1 hour, and FY-4A / B satellite remote sensing data is 15 minutes. The airborne detection time series data is processed for 5 minutes, and the midpoint of the window is used as the representative time. ERA5 and FY satellites use temporal nearest neighbor matching to merge these features into a unified feature vector, which is then associated with the icing intensity label to form a complete training sample. The above process is repeated for all windows of all flights, and all generated samples are concatenated to finally output a structured icing sample dataset (CSV file or DataFrame).

[0043] Step S102: Train and test multiple base models using the ice accumulation sample dataset, and obtain the prediction accuracy of each base model for different ice accumulation levels based on the confusion matrix.

[0044] As one implementation method, multiple base models are trained and tested using an icing sample dataset. The prediction accuracy of each base model for different icing levels is obtained based on the confusion matrix, including: The icing sample dataset is divided into a training set and a test set. Each base model is trained using the training set. The predicted level labels of each base model are obtained using the test set. Based on the level labels and predicted level labels of each base model, a confusion matrix is ​​constructed to obtain the prediction accuracy of each base model for different icing intensity levels.

[0045] In some embodiments, the division criteria for the training set and the test set are based on the time of flight sorties. All data from each flight mission are treated as a whole. For example, if the sample is airborne detection data from 2021 to 2025, 80% of the sorties from each year's flight sortie set are randomly selected for the training set and 20% for the test set, ensuring that both the training set and the test set cover flight sortie samples from each year from 2021 to 2025. The representativeness of the spatial distribution is checked by viewing the spatial distribution of sample points in the training set and the test set on a map, and checking whether the test set covers different geographical sub-regions, such as plains, mountains, and coastlines, to ensure that there is no obvious spatial bias. The balance of the category distribution is checked by statistically analyzing the percentages of "no, light, medium, and heavy" icing levels in the training set and the test set. If the test set is severely lacking in "moderate" or "severe" icing samples, flights containing these rare events should be manually assigned to the test set to ensure that the model can be effectively evaluated. Window randomization involves randomly shuffling the 5-minute windows within each flight after determining which flights are included in the training and test sets, to prevent the model from learning time-series patterns within flights and to allow it to focus more on learning the causal relationship between weather conditions and icing.

[0046] As one implementation approach, base models include Random Forest, XGBoost, and LightGBM.

[0047] Random Forest (RF), XGBoost (XGB), and LightGBM (LGB) were selected as the base models. These three models can effectively handle nonlinear relationships and high-dimensional features, but they each have their own strengths in algorithm implementation, efficiency, and handling of data characteristics, providing a foundation for subsequent performance differentiation analysis and ensemble. Furthermore, the core tuning parameters and strategies for these three training models are shown in Table 2.

[0048] Table 2 Core tuning parameters and strategies for the three models

[0049] As shown in Table 2, for Random Forest, setting `class_weight='balanced'` or using custom weights forces the model to focus more on the few but dangerous heavily icing categories. For XGBoost, adjusting `scale_pos_weight` (for a simplified approach to binary classification) or multi-class sample weights prioritizes the ability to capture moderate and heavy icing (i.e., recall). For LightGBM, class imbalance is addressed using `class_weight` or `is_unbalance` parameters. The hyperparameters of each model (n_estimators, max_depth for RF; learning_rate, max_depth, and regularization parameters for XGB; num_leaves, learning_rate, etc. for LGB) have all been optimized to improve the model's ability to identify moderate and heavy icing while preventing overfitting.

[0050] The prediction results of each base model are obtained using the test set. Based on the prediction results and the true labels, a confusion matrix containing four categories—no ice accumulation, light ice accumulation, moderate ice accumulation, and heavy ice accumulation—is calculated for each base model. To objectively evaluate and construct the confusion matrix, the remaining 20% ​​of the ice accumulation sample dataset is used as the test set. The test set is input into the three trained models to obtain their corresponding prediction results y_pred_M, where M is the model number (M... The prediction results for the three samples {RF, XGB, LGB} are y_pred_rf, y_pred_xgb, and y_pred_lgb, respectively. For each model M, the prediction result y_pred_M on the test set is compared with the true icing label y_true, and a 4×4 confusion matrix (CMM) is calculated. The rows of the matrix represent the true class, and the columns represent the predicted class. The order of the four classes is fixed as: ['No Icing', 'Light Icing', 'Moderate Icing', 'Heavy Icing']. The prediction accuracy is calculated based on the confusion matrix corresponding to each base model. The prediction accuracy is the proportion of samples that the model correctly predicts as that level. For model M and a specific icing level j (e.g., j='Heavy Icing'), its prediction accuracy is... The calculation formula is: , in, It is the total number of samples predicted as level j in model M (the sum of this column). This represents the number of samples correctly predicted as rank j by model M (values ​​on the diagonal of the confusion matrix). Prediction accuracy. It measures the model's reliability in predicting at a certain level. Reliability is not the probability value or confidence score output by the model in real time, but rather the historical prediction accuracy of a model for a specific level. This is a global, post-hoc statistical indicator calculated from the confusion matrix of the independent test set.

[0051] S103. For each ice accumulation level, select the base model with the highest prediction accuracy as the expert model for that ice accumulation level, and obtain the expert model lookup table.

[0052] As one implementation method, for each of the four icing levels j (no icing, light icing, moderate icing, and heavy icing), the prediction accuracy of the three base models at that level is compared. Includes three prediction accuracy rates , , For each level j, select the prediction accuracy. The highest-level base model is designated as the "expert model" for that level, ultimately forming a structured lookup table (or mapping dictionary), as shown in Table 3.

[0053] Table 3 Expert Model Lookup Table

[0054] As can be seen from the above statement, its core logic is not to blindly trust all the outputs of any single model, but only to trust the outputs of the specific areas in which it has performed best and most reliably in historical testing. Through the expert model lookup table, the system "assigns" the forecast tasks of different ice accumulation levels to the specific model that is most reliable in the historical records for that level.

[0055] The expert model lookup table can be periodically re-evaluated and updated based on newly added probe data to achieve continuous model optimization. Specific update strategies can be configured according to business needs. Prioritizing system stability, after a certain period of accumulation (e.g., quarterly or annually), at the end of the cycle, the newly accumulated data is used as a new independent test set. Predictions are made using the existing integrated model and compared with the prediction results of the current business system. If the accuracy difference is controlled within ±5%, the system is considered stable and no update is needed. If a model's new accuracy at a certain critical level is significantly and consistently higher than the current expert model (e.g., more than 10%), an update is triggered. After the updated expert model lookup table is reviewed and deemed reasonable by domain experts, it will be deployed to the business system as a new version, while the old version is retained for rollback.

[0056] S104. Obtain forecast field data, construct a pattern feature dataset based on the target flight path, input the pattern feature dataset into multiple trained base models, and obtain preliminary icing level prediction results.

[0057] As one implementation method, forecast field data is acquired, and a pattern feature dataset is constructed based on the target flight path, including: Acquire operational numerical model forecast field data with a spatial resolution of hundreds of meters and a temporal resolution of minutes; extract meteorological elements from the forecast field data that are consistent with the type of the icing sample dataset to obtain a three-dimensional meteorological element field; determine the target flight route, which consists of multiple waypoints arranged in time sequence, including time, longitude, latitude, and altitude information for each waypoint; extract meteorological feature vectors for each waypoint based on the three-dimensional meteorological element field; arrange the feature vectors of all waypoints in order according to the target flight route to obtain the model feature dataset.

[0058] In some embodiments, forecast field data products from operational high-precision numerical weather prediction models with horizontal resolution at the hundred-meter level and temporal resolution at the minute level are used as the data source. From the forecast field data, a three-dimensional meteorological element field is extracted that is completely consistent with the meteorological elements used in the training samples. This includes elements corresponding to airborne sounding data, derived features calculated from elements corresponding to ERA5 reanalysis data, and elements corresponding to the roles in satellite remote sensing data. The target flight path is determined, i.e., the target flight path to be predicted, and it is specified that each waypoint must contain four core attributes: forecast time, longitude, latitude, and flight altitude. This information collectively determines the spatial location where the aircraft will arrive at a future time.

[0059] As one implementation method, meteorological feature vectors for each waypoint are extracted based on a three-dimensional meteorological element field, including: The system locates waypoints based on the timestamp, longitude, latitude, and altitude of a 3D meteorological field. It then performs spatial interpolation of the waypoint's longitude and latitude within the 3D meteorological field to obtain the corresponding terrain altitude. The system calculates the relative ground altitude based on the altitude and terrain altitude. Based on the relative ground altitude, it estimates the waypoint pressure using static equations and normalizes it to Sigma layer coordinates in the terrain-following coordinate system, resulting in a waypoint time series containing the timestamp, longitude, latitude, and Sigma layer coordinates. Finally, it selects time data corresponding to the timestamp, obtains horizontal meteorological element values ​​using bilinear interpolation based on longitude and latitude, and obtains vertical meteorological element values ​​using linear interpolation based on the Sigma layer coordinates. Finally, it generates the meteorological feature vector based on the time data, horizontal meteorological element values, and vertical meteorological element values.

[0060] The time, longitude, latitude, and altitude of waypoints are located in a three-dimensional meteorological element field. All required meteorological element values ​​corresponding to the waypoints are extracted using spatiotemporal matching and interpolation methods. Considering the requirements of low-altitude flight and artificial rain enhancement flights, the flight path altitude data is expressed as altitude, but terrain following needs to be taken into account. The time, longitude, latitude, and altitude H of the waypoints are read. msl_i Extract the surface topography field corresponding to the latitude and longitude of waypoints from the three-dimensional meteorological element field, and obtain the topography value H of each waypoint using nearest neighbor or bilinear interpolation. terrain_i The relative ground elevation H of each waypoint is obtained by subtracting the terrain value from the altitude. agi_i The pressure at this height can be approximated using static equations: Among them, P i P is the air pressure at the waypoint. sfc_i Let g be the ground air pressure at the waypoint, g = 9.81 m / s², R = 287 J / kg·K, and T be the gas constant for dry air. avg_i This represents the near-surface mean temperature of the model. Elevation is converted to the sigma layer (σ).i ): , Among them, P top The top pressure of the model is typically 50 hPa, σ i The value range is 0-1. The time sequence for obtaining the waypoints is time T. i Longitude Lon i Latitude i σ layer i Spatiotemporal interpolation: Since the time resolution of both the flight path and the 3D meteorological element field is 1 minute, no time interpolation is required. Horizontal spatial interpolation: For a sigma layer (Lon... i Lat i Find the bounding 4-grid points (Lon0, Lat0) to (Lon1, Lat1) and perform bilinear interpolation. Vertical interpolation: find the bounding σ... i Adjacent layer σ low >σ i >σ high Linear interpolation is performed using adjacent layers.

[0061] The meteorological element values ​​extracted from all waypoints are organized and assembled, with all element values ​​extracted from each waypoint forming a feature vector, which serves as an input sample for the model. The feature vectors of all waypoints are arranged in order of the flight path, ultimately forming a structured model feature dataset. In this dataset, each row represents a waypoint, and each column represents a meteorological element feature of an input model. Its data format and feature dimensions are completely consistent with the icing sample dataset.

[0062] As one implementation method, the pattern feature dataset is input into multiple trained base models to obtain preliminary icing level prediction results, including:

[0063] For each waypoint, multiple icing level predictions are obtained based on several trained base models. The pattern feature dataset, or samples from each waypoint sequentially, are input into the three trained base models: Random Forest (RF), XGBoost (XGB), and LightGBM (LGB). These three models have been optimized using historical data and possess the ability to map meteorological features to icing levels. Each model determines the icing level for each waypoint along the entire route based on its internally learned rules. Model RF calculates the predicted icing level for each sample, denoted as y_pred_rf. Model XGB calculates the predicted icing level for each sample, denoted as y_pred_xgb. Model LGB calculates the predicted icing level for each sample, denoted as y_pred_lgb. After the above process is completed, for each waypoint on the target route, three possible or different icing level prediction results will be obtained (y_pred_rf[i], y_pred_xgb[i], y_pred_lgb[i], where i represents the i-th waypoint). These three prediction results together constitute the preliminary icing level prediction results of each base model.

[0064] S105. Based on the preliminary icing level prediction results, the final icing level prediction results of the target flight path are obtained by using an expert model lookup table.

[0065] As one implementation method, the preliminary icing level predictions (y_pred_rf, y_pred_xgb, y_pred_lgb) of the three base models (RF, XGB, LGB) corresponding to each waypoint on the target route are used as predictions to be decided. An expert model lookup table is used as the decision rule base. Table 3 defines the expert model with the highest historical prediction accuracy for each icing level (none, light, moderate, severe). Based on the preliminary prediction level results of the three models, the expert model lookup table (Table 3) is queried to determine the historical reliability of each model in the current prediction for this specific level.

[0066] For example: if the RF predicts "moderate icing," then looking up the table, the historical accuracy rate of RF as a "moderate icing expert" is 0.82, denoted as conf_rf; if the XGB predicts "severe icing," then looking up the table, the historical accuracy rate of XGB as a "severe icing expert" is 0.85, denoted as conf_xgb; if the LGB predicts "no icing," then looking up the table, the historical accuracy rate of LGB as a "no icing expert" is 0.98, denoted as conf_lgb. Then, the confidence levels are compared and the best value is selected. That is, the three confidence values ​​(conf_rf, conf_xgb, conf_lgb) are compared, and the prediction result of the base model corresponding to the highest confidence value is selected as the final icing target prediction level for that waypoint. As shown in the previous example, conf_lgb (0.98) > conf_xgb (0.85) > conf_rf (0.82). Therefore, the system adopts the prediction of the LGB model and finally outputs "no icing". After traversing all waypoints, a sequence composed of the final forecast level of each point is obtained. This sequence is the final icing level prediction result of the target flight route for this specific target route after dynamic model integration optimization.

[0067] S106. Based on the final icing level prediction results of all waypoints of the target flight route, generate a visualized icing potential forecast product. The visualized icing potential forecast product includes: using geographic information as a base map, drawing the target flight route as a continuous curve, and marking the corresponding waypoint locations with different colors or graphic markers according to the final icing level prediction results of each waypoint.

[0068] As one implementation method, the final icing level prediction result is a structured list of forecast results. This list clearly defines the final forecast status of each waypoint on the target flight path, including at least the longitude, latitude, forecast time, flight altitude, and corresponding final icing target forecast level (e.g., none, light, moderate, heavy) for each point. A base map and flight path are then drawn, using a geographic information map containing topography, airspace, and important landmarks as the background map. On this base map, the target flight path is drawn as a continuous curve or a path composed of line segments based on the latitude and longitude coordinates of the waypoint sequence. Icing risk levels are then labeled. Based on the forecast result list, along the drawn flight path, the final icing level is visually coded at the spatial location corresponding to each waypoint.

[0069] Color coding systems are typically used: for example, Figure 2-3As shown, "green" marks "no icing" segments; "blue" marks "light icing"; "orange" marks "moderate icing"; and "red" marks "heavy icing." Colors can be directly filled into the corresponding segments of the flight path, or marked as color blocks or dots at waypoint locations. Graphical or symbolic markings can also be used, such as using different shaped icons (e.g., triangles, exclamation marks) at waypoints of different icing levels for emphasis. Finally, add necessary legends (explaining the correspondence between colors / symbols and icing levels), scale bars, flight path markings, forecast validity periods, and other auxiliary information to form a complete thematic map. Figure 2 This is a two-dimensional schematic diagram of the visualized ice accumulation potential forecast product provided in this application example. Figure 3 A three-dimensional schematic diagram of the visualized ice potential forecast product provided as an example in this application.

[0070] As the above process demonstrates, this method is targeted; the product is not a generalized weather map covering a wide area, but a customized product specifically generated for a pre-defined flight path. It also intuitively condenses complex multidimensional data and model calculation results into a flight path with clearly defined color-coded risk indicators. Pilots, flight dispatchers, or weather modification operation commanders can grasp the risk distribution of the entire flight path at a glance—where it is safe and what degree of icing threat exists—without needing to interpret professional meteorological isolines. Moreover, this intuitive format provides direct evidence for real-time flight decision-making, enabling the planning of safe routes, execution of emergency evasive maneuvers, or assessment of the flight safety window for weather modification aircraft. This achieves seamless integration between forecast products and user decision-making scenarios, greatly improving the practicality and operational efficiency of forecasts.

[0071] This application provides a method, device, electronic equipment, and storage medium for flight path icing forecasting based on dynamic integration of multi-source data and models. By deeply fusing multi-source data, it constructs an icing sample dataset encompassing macroscopic atmospheric background, cloud structure, and microscopic cloud physical states, providing a rich data foundation for model training. Based on a confusion matrix, it dynamically integrates and optimizes multiple models, and through in-depth analysis of the performance characteristics of different machine learning models in forecasting different levels of icing, it constructs a dynamic and reliable integrated forecasting system, significantly improving the overall accuracy of icing forecast levels. Employing a high-resolution numerical model, it performs precise spatiotemporal data mining and interpolation on the target flight path, transforming the gridded meteorological forecast field into a sequential feature dataset that is isomorphic to historical training data along a predetermined trajectory. This enables the direct conversion of grid point forecasts into customized, high-resolution numerical icing risk forecasts for specific flight paths. Finally, it transforms the digitized icing forecast results into intuitive, easy-to-understand, and directly serviceable graphical business products for flight decision-making, providing direct evidence for real-time flight decisions and achieving seamless integration between forecast products and user decision-making scenarios, greatly enhancing the practicality and operational efficiency of forecasts. This method effectively overcomes the shortcomings of traditional forecast results being generalized and crude, significantly improving the reliability of identifying moderate and severe hazardous icing and the overall forecast accuracy. It provides precise, intuitive, and operable key technical support for pilots to avoid risks, ensure low-altitude flight safety, and safely schedule aircraft for weather modification operations, and has high operational application value and socio-economic benefits.

[0072] To achieve the above embodiments, this application also proposes a route icing forecasting device based on dynamic integration of multi-source data and models. Figure 4 This is a schematic diagram of a route icing forecasting device based on dynamic integration of multi-source data and models, provided as an embodiment of this application. Figure 4 As shown, the route icing forecasting device based on multi-source data and dynamic model integration may include: a data acquisition module 401, a model training module 402, a model selection module 403, a model application module 404, and an icing prediction module 405.

[0073] Among them, the data acquisition module 401 is used to acquire multi-source data to construct an ice accumulation sample dataset. The multi-source data includes airborne detection time series data, reanalysis data and satellite remote sensing data. The model training module 402 is used to train and test multiple base models using an ice accumulation sample dataset, and to obtain the prediction accuracy of each base model for different ice accumulation levels based on the confusion matrix. The model selection module 403 is used to select the base model with the highest prediction accuracy for each ice accumulation level as the expert model for that ice accumulation level, and obtain the expert model lookup table. Model application module 404 acquires forecast field data, constructs a pattern feature dataset based on the target flight path, inputs the pattern feature dataset into multiple trained base models, and obtains preliminary icing level prediction results. The icing prediction module 405 uses an expert model lookup table based on the preliminary icing level prediction results to obtain the final icing level prediction results for the target flight path.

[0074] In some implementations, the apparatus also includes a model product generation module 406, used for: Based on the final icing level prediction results of all waypoints along the target flight route, a visualized icing potential forecast product is generated.

[0075] It should be noted that the foregoing explanation of the embodiment of the route icing forecasting method based on dynamic integration of multi-source data and models also applies to the route icing forecasting device based on dynamic integration of multi-source data and models in this embodiment, and will not be repeated here.

[0076] To implement the above embodiments, this application also proposes an electronic device. Please see [link to relevant documentation]. Figure 5 , Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 5 As shown, the electronic device 500 includes: a processor 501 and a memory 502 communicatively connected to the processor 501; the memory 502 stores computer-executable instructions; the processor 501 executes the computer-executable instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0077] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0078] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0079] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0080] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0081] This application is intended to provide an implementation scheme for users to selectively prevent the use or access to their personal information data. Specifically, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.

[0082] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0083] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0084] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0085] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0086] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0087] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0088] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0089] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for forecasting airway icing based on dynamic integration of multi-source data and models, characterized in that, Includes the following steps: A dataset of ice accumulation samples was constructed by acquiring multi-source data, including airborne detection time-series data, reanalysis data, and satellite remote sensing data. Multiple base models were trained and tested using the aforementioned ice accumulation sample dataset, and the prediction accuracy of each base model for different ice accumulation levels was obtained based on the confusion matrix. For each ice accumulation level, the base model with the highest prediction accuracy is selected as the expert model for that ice accumulation level, and an expert model lookup table is obtained. Acquire forecast field data, construct a pattern feature dataset based on the target flight path, input the pattern feature dataset into multiple trained base models, and obtain preliminary icing level prediction results; Based on the preliminary icing level prediction results, the final icing level prediction results for the target flight path are obtained using an expert model lookup table.

2. The method according to claim 1, characterized in that, The airborne detection time series data includes icing intensity, temperature, relative humidity, dew point temperature, liquid water content, cloud droplet number concentration, effective cloud droplet diameter, and atmospheric precipitable water; the reanalysis data includes vertical velocity, inversion layer intensity, and water vapor flux; the satellite remote sensing data includes cloud top height, cloud top temperature, supercooled layer thickness, and cloud optical thickness.

3. The method according to claim 2, characterized in that, The process of acquiring multi-source data to construct an ice accumulation sample dataset includes: The airborne detection time series data is acquired, and the airborne detection time series data is aggregated based on a preset time window and sliding step size to obtain the sample time, spatial location and feature vector of each time window; The ice accumulation intensity is divided into four ice accumulation intensity levels, and the ice accumulation intensity level with the highest frequency in each time window is used as the level label for that time window. Based on the alignment of the sample time and spatial location, the corresponding reanalysis data and satellite remote sensing data are extracted to obtain derived features and satellite regional features, respectively. The time, spatial location, feature vector, derived features, satellite region features, and grade labels of the samples are fused to form a single sample, and all the single samples constitute the ice accumulation sample dataset.

4. The method according to claim 3, characterized in that, The process of training and testing multiple base models using the icing sample dataset, and obtaining the prediction accuracy of each base model for different icing levels based on the confusion matrix, includes: The ice accumulation sample dataset is divided into a training set and a test set; Each base model is trained using the training set. The prediction level labels of each base model are obtained using the test set; Based on the level labels and predicted level labels of each base model, a confusion matrix is ​​constructed to obtain the prediction accuracy of each base model for different ice intensity levels.

5. The method according to claim 1, characterized in that, The base models include Random Forest, XGBoost, and LightGBM.

6. The method according to claim 1, characterized in that, The acquisition of forecast field data, based on the construction of a pattern feature dataset according to the target flight path, includes: Acquire operational numerical model forecast field data with spatial resolution at the hundred-meter level and temporal resolution at the minute level; Extract meteorological elements from the forecast field data that are of the same type as the ice accumulation sample dataset to obtain a three-dimensional meteorological element field; Determine the target flight route, which consists of multiple waypoints arranged in a time sequence. The waypoints include time, longitude, latitude, and altitude information. Meteorological feature vectors for each waypoint are extracted based on the three-dimensional meteorological element field. The feature vectors of all waypoints are arranged in order according to the target flight route to obtain the pattern feature dataset.

7. The method according to claim 6, characterized in that, The extraction of meteorological feature vectors for each waypoint based on the three-dimensional meteorological element field includes: Based on the timestamps, longitudes, latitudes, and altitudes of the waypoints located in the three-dimensional meteorological field; Based on the longitude and latitude of the waypoints, spatial interpolation is performed in the three-dimensional meteorological element field to obtain the corresponding terrain height; Calculate the relative ground height based on the altitude and terrain elevation; Based on the relative ground height, the waypoint pressure value is estimated using the static equation and normalized to the Sigma layer coordinate value of the terrain-following coordinate system to obtain the waypoint time series including timestamp, longitude, latitude and Sigma layer coordinate value; Select the time data corresponding to the timestamp, obtain the horizontal meteorological element values ​​based on the longitude and latitude using bilinear interpolation, obtain the vertical meteorological element values ​​based on the Sigma layer coordinate values ​​using linear interpolation, and generate the meteorological feature vector based on the time data, horizontal meteorological element values ​​and vertical meteorological element values.

8. The method according to claim 1, characterized in that, The preliminary icing level prediction results include: For each waypoint, multiple icing level predictions are obtained based on multiple trained base models.

9. The method according to claim 1, characterized in that, After obtaining the final icing level prediction result of the target flight path based on the preliminary icing level prediction result using an expert model lookup table, the method further includes: Based on the final icing level prediction results of all waypoints of the target flight route, a visualized icing potential forecast product is generated; the visualized icing potential forecast product includes: using geographic information as a base map, drawing the target flight route as a continuous curve, and marking the corresponding waypoint locations with different colors or graphic markers according to the final icing level prediction results of each waypoint.

10. A route icing forecasting device based on dynamic integration of multi-source data and models, characterized in that, include: The data acquisition module is used to acquire multi-source data to construct an ice accumulation sample dataset, wherein the multi-source data includes airborne detection time series data, reanalysis data, and satellite remote sensing data; The model training module is used to train and test multiple base models using the ice accumulation sample dataset, and to obtain the prediction accuracy of each base model for different ice accumulation levels based on the confusion matrix. The model selection module is used to select the base model with the highest prediction accuracy for each ice accumulation level as the expert model for that ice accumulation level, and obtain the expert model lookup table. The model application module acquires forecast field data, constructs a pattern feature dataset based on the target flight path, and inputs the pattern feature dataset into multiple trained base models to obtain preliminary icing level prediction results. The icing prediction module uses an expert model lookup table based on the preliminary icing level prediction results to obtain the final icing level prediction results for the target flight path.

Images